diff --git "a/measurement.json" "b/measurement.json"
new file mode 100644--- /dev/null
+++ "b/measurement.json"
@@ -0,0 +1,62439 @@
+{
+ "measurement": {
+ "model.layers.0.self_attn": [
+ {
+ "accuracy": 0.8348713423076429,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8563541111193205,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8810938910434121,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9182686993950292,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9216420148548327,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9212327693638048,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9400708738126253,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9403981660541735,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9484252176786724,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9506724319959942,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.959431391013296,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9636056078107733,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9638395717269496,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9685677038995844,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9795099920348117,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836703742805281,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9826040832619918,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9916155989232817,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9956198146468714,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.0.mlp": [
+ {
+ "accuracy": 0.7990592404415733,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8049562353836863,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.837525907315706,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.848137617111206,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8997046821995786,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9079057229192633,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.922072780759711,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9472709517729909,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9524140891275907,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9489443145300213,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9553703038316024,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.973903521111137,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9777101560642845,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9856718574699602,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9866912584555777,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9898408376856854,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9960923622313299,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.1.self_attn": [
+ {
+ "accuracy": 0.8466421302996183,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8579504113448293,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8708957621925756,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9178714501230341,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9236355706265098,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9241353963550768,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9474732593486184,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9483000385133844,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9534335920685216,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9562876130405226,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9608239060954044,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9633180561818575,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9650913571056566,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9679142173967863,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9807046291075254,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836615386762118,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9856140393959848,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.991104856133461,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9961273885871235,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.1.mlp": [
+ {
+ "accuracy": 0.9121535326305189,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.944867121545892,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9451748132705688,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.946027253803454,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9790729441140827,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9844773741144883,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986789904142681,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9904079923504278,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922495782375336,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9914402553909704,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9927847252080315,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9955984326569658,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.996232334720461,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9972205636532683,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977208589644808,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.998199992862187,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.998939460045413,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.2.self_attn": [
+ {
+ "accuracy": 0.980812342543351,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9817478703825098,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836911198339964,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9865352605518541,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9891414948199925,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887690410802239,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9925083808208767,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9923683020629381,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930588508907118,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933103824916639,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933857572706122,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937914562852759,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9951649555250218,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.995419935176247,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9972783338866735,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.997722710825895,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977120425748197,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.998962314799428,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.999371977915105,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.2.mlp": [
+ {
+ "accuracy": 0.9592505505210475,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9604840686446742,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9661218837687844,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9678960630768224,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9793339654019004,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9809346481373435,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9834233271448236,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9891691270627474,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9901300402064073,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.98946528999429,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9907386200992685,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9946373093285059,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9953858593576833,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9970687065077456,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9972699484542796,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977825127149883,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.999210785878332,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.3.self_attn": [
+ {
+ "accuracy": 0.9591080295412164,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9607760874848617,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9667377064102575,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9707765673336229,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9791299292915746,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9802252242439672,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9888263893754858,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9886720611860877,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9897664361878445,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9900349692294472,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9897192622485914,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9902863455446143,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928861511381049,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9932500649439661,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9959878090180849,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9966145326432428,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9967585268773531,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.998549910458295,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9991777357400248,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.3.mlp": [
+ {
+ "accuracy": 0.9451740478214464,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9468001315468236,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9546740023713363,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9570689515063637,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9721979467492354,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9743350948158064,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9777054363175443,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9855049111341175,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867324264425981,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9858475264750028,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9875501260945672,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928098462129894,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9938018741576296,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9960847128378717,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9963409779103178,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9970253880478834,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9989670489571596,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.4.self_attn": [
+ {
+ "accuracy": 0.9505745643063596,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9555195632733797,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9603777653292606,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9672843249220597,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9747963547706604,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9754832653622878,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9851507183752561,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9859509515134912,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867505148837441,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9872978133590598,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9871106116395247,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9880341118887851,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9906246630769027,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9913984003819918,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9948302656412125,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9955168687983563,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9960200919916755,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9979499531419653,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9989579461122814,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.4.mlp": [
+ {
+ "accuracy": 0.930737620905826,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9329659060428017,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9438573498474925,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9470321630176745,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9651784489029333,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9679425201917949,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9723987516603971,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9818342673151117,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9833689677087885,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9822716148276078,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9844477490374917,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9909824274088207,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922494864777515,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9950621677072424,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9954056081018949,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9963421915706835,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.99868955463171,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.5.self_attn": [
+ {
+ "accuracy": 0.9477525635769493,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9500151496184499,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9581744294417531,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9652545985422636,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9747992299104992,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9750199223819532,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9854097303591276,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9858022143966273,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9865916277232923,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9871616504694286,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867476005303232,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9879800410647142,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9902426459287342,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9909198080238543,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.994668358642804,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9955465334810709,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9960868401747001,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9976765011486254,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9989541977840034,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.5.mlp": [
+ {
+ "accuracy": 0.9189943326146979,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9216764537911666,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9343216230994776,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9379867064325433,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9593618167074103,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9625389795554311,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9677558698152241,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9788212807554948,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9805729593101301,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9793009303118053,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9818327238685206,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9894699822915228,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.990947309293245,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9942310323056421,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9946357984291879,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9957197351675284,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9984681782753844,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.6.self_attn": [
+ {
+ "accuracy": 0.940226159597698,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.942072447977568,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9511398359348899,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.96212006870069,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9704323379616988,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9708967899021349,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9838833746157194,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9843400277589497,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9855711397371794,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986207863217906,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9854968252934908,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9864261667979392,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.988646114343091,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9894326548827322,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937960697632087,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9947181249919691,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9958227214060331,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9973396294211087,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9988965125460374,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.6.mlp": [
+ {
+ "accuracy": 0.9106139697526631,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9133718641180741,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9276771670893619,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9317521358791151,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9550079170026278,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.958581877382178,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9644253002969843,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9766399091795871,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.978563583210895,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9770508486973611,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9798655651117626,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9883031868620923,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9899602728454691,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9936309480353406,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940370235003924,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.995271550981622,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9983208751992175,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.7.self_attn": [
+ {
+ "accuracy": 0.9353753516548559,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9369817783958033,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9457551868338334,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9585292841258802,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9671356113333451,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9674881728071916,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9816739449375554,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9822429685216201,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836901645911368,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9845326887933832,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836596360332087,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.984762977612646,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9868164564433851,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9878586138549604,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928263480726042,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9939228582539057,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9953555053001956,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9968355219615134,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987706711613819,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.7.mlp": [
+ {
+ "accuracy": 0.9060384035110474,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9088548233634547,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9241834376987658,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9285457698922408,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9524270766659787,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9563243326387907,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9625766277313232,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9752391984588221,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9773524152605158,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9756978216924166,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9787147421585886,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9876035106809515,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9893762892798373,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9932421853667811,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9936758844475997,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9950494781920785,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9982252966024374,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.8.self_attn": [
+ {
+ "accuracy": 0.9283549346421894,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9310015502728914,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9395622642416703,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9544188380241394,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9635292323012101,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9638580585780897,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9794533080176303,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9799409351850811,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9815086298867276,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9824706598332054,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9818307286814639,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9830946420368395,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9850397015872755,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9862222248002103,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9918626413533562,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930871638812517,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9947133460327199,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9963636131663072,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985840627433438,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.8.mlp": [
+ {
+ "accuracy": 0.9035096482226723,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9064070237310309,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9216761400825099,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.926001460928666,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9511726184895164,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9551418611877843,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.961395684041475,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.974639649453916,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9767268896102905,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9750217152269263,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9781331529742793,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9872406603474366,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9890861879838141,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930410612570612,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9934701507813052,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9948159126858962,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9981612370986688,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.9.self_attn": [
+ {
+ "accuracy": 0.9205297608124583,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9255120691500212,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9345132802662096,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9483281248494199,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9601302962554128,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9601875010289644,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9773775574408079,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9774408591421027,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9794017547055295,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.98048744703594,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9795870843686556,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9805500601467333,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9831974788715965,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.98491128965428,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9908731383712668,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9924184711355912,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937756245857791,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.996222795624482,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9984297538666349,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.9.mlp": [
+ {
+ "accuracy": 0.9016373847660265,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9045906129636263,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9199933503803454,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9243862377969843,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9502467387600949,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9542844766064694,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9606036010541414,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9741416949974863,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9762628203944156,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9745790707437616,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9777181430866844,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9870286921137258,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9888802766799927,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929252133557671,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933747108045378,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9947460215342673,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9981288268770042,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.10.self_attn": [
+ {
+ "accuracy": 0.9241868320264315,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9278458105890375,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.936782673785561,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9518021094171625,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9619670384808591,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9618789961463526,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9784853238808481,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9788413345813751,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9804527069392958,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9815383412336048,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9805720260268763,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9818528906295174,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9841732963135368,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9854076363538441,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9914044400579051,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9926862810787401,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943352157348081,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9960806620748419,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985131208637827,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.10.mlp": [
+ {
+ "accuracy": 0.90056522268998,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.903525239542911,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9190181807467812,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9234152530369005,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.949757594811289,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9537839701301173,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9601684400909826,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9740112941516074,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9761010000580236,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9743614808509224,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9775130466410988,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986932267483912,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887840943901163,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929048073919196,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933267596520876,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9946672983075443,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9981265332745878,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.11.self_attn": [
+ {
+ "accuracy": 0.9245347474750719,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9270116592708387,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9363818105898405,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9498730239115263,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9626116501657587,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9628933479911402,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9789095128837385,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9793721939388075,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9809632160161671,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.981840979111822,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9810076983351457,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9823228277658161,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9843895435333252,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9856174462719968,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9916104313574339,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9927186785559905,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9945593562565351,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9961846694350243,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985444859454506,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.11.mlp": [
+ {
+ "accuracy": 0.8989769596802561,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9020528542368036,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9177310341282895,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9222558736801147,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.948719670898036,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9529156339795966,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9593884631207115,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9732893956334967,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9754655345490104,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9737294535887869,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9769870058486336,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9865649891527075,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9884947514847705,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9926635097516211,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.993141150788257,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9945505092802801,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.998068978911952,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.12.self_attn": [
+ {
+ "accuracy": 0.9175767082917063,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.920829082790174,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9335885424362986,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9505112390769155,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9590961242976942,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.959463894367218,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9771202112499037,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9773993711722525,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9791853208290903,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9801424882913891,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9795772056830557,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9806756894839438,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836236395333943,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9848008344047948,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9910853144369627,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9924198028288389,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940394061176401,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9959663623257687,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9983953712998253,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.12.mlp": [
+ {
+ "accuracy": 0.89693284034729,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9000938252398842,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9162795857379311,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9210276039023149,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9475858525225991,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9519705772399902,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9586295391383924,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9723975407449823,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9747448108698192,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9731564929610804,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9765334866548839,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9862599153267709,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9882604675857645,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9923921462736631,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929963562049364,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.994492670815242,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.997979932905812,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.13.self_attn": [
+ {
+ "accuracy": 0.9124533565420854,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9164229505940488,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.927198052406311,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9429556884263691,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9553787488686412,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9561712334030553,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9756978044384405,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9761855539522672,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9779441278231772,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9791419004139147,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9780364177728954,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9796405011101773,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9820030551207694,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.983441804584704,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.990430706425717,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9917423050654562,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9936587818359074,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9956773797932424,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9983236887737325,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.13.mlp": [
+ {
+ "accuracy": 0.8936305422531932,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8969561489004838,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9136605701948467,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9185545883680645,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9461449196464137,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9505808761245326,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9574435510133442,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9717390035328112,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9741114111323106,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9724815142782111,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9759268101892973,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9859460435415569,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9879805563311828,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922434985637665,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928472410691412,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943253805762843,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9979344135836551,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.14.self_attn": [
+ {
+ "accuracy": 0.9128959555374949,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9160991467927632,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9264013641758969,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9433889012587697,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.956124628845014,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9567425564715737,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9745036755737505,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9752163918394792,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9770422358261911,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.978490162836878,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9780091690389734,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9794250209080545,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9816573685721347,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9831443852499911,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9899518819231736,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.991578440132894,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9934366749305474,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9953809560913789,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9982669945610197,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.14.mlp": [
+ {
+ "accuracy": 0.8873454897027266,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8909815110658345,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9086400395945499,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9140274587430453,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9429366776817724,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9476562550193385,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9549706013579118,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9698275955099809,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9723402638184396,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9708241130176344,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9744876576097388,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9851011241737165,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9872511507649171,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9916999543968,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.992428144342021,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940195742406344,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977901199538457,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.15.self_attn": [
+ {
+ "accuracy": 0.9098503589630127,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9102449103405601,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9228094377015766,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9396675134959974,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.952316657492989,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9539488114808735,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9731182299162212,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9745106869622281,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9760193118923589,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9774818310612127,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9769615201573623,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9785761550853127,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9810286305452648,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9826658258312627,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9895090288237521,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9913618823415354,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929540941589757,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9955948253995494,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9981213412002513,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.15.mlp": [
+ {
+ "accuracy": 0.8824305157912404,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8862165526339882,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.90499284392909,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9106200055072182,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9407878863184076,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9455742271322953,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.953282538213228,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9688537654123808,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9713983096574482,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.969781452103665,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9735053150277388,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.984584706394296,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867760633167467,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9914684507407641,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9921577078731436,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.993803399173837,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977195986399525,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.16.self_attn": [
+ {
+ "accuracy": 0.9211047511351735,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9242972888444599,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9342741778022364,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9496132009907773,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9599781946132058,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9610964718617891,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9756880590790197,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9770794495155937,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9785390245287042,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9795664909638857,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9801185742804879,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9812341677515131,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9834693262451574,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9848615354613254,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9908077363905153,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9924038930943138,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9934591372546396,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9959034762884441,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9983107059409744,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.16.mlp": [
+ {
+ "accuracy": 0.8840434425755551,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8876348043742933,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9059883607061285,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9116571576971757,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9414021153199046,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9461048151317396,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9538153817779139,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9692259744593972,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9717851814470793,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9700904927755657,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9737874253800041,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9847466945648193,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869211551390196,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9915795271333895,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922418413977874,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.993868132170878,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977417788223216,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.17.self_attn": [
+ {
+ "accuracy": 0.9279488952536332,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9310202849538702,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9415805402554964,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9543592647502297,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9634423946079455,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9644073373393008,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9781844223800459,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9794418874539828,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9807027797949942,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9816716219249525,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9813549847979295,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9825067473085303,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.985111813796194,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986191263324336,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9917667445383573,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930872783849114,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943026912055517,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9965091994718501,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9984704274684191,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.17.mlp": [
+ {
+ "accuracy": 0.8836462246744257,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8871662240279348,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9059722611778661,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9116674285185964,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9412784450932553,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9459992333462364,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9538214896854601,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9691127381826702,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.971726728113074,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9700328864549336,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9737456770319688,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.984702119701787,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9868926915683245,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9915344770017424,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922341367131785,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9938796942955569,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977270521615681,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.18.self_attn": [
+ {
+ "accuracy": 0.9409106091449135,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9420342257148341,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9503616157330965,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9600670620014793,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9691581443736428,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9709192106598302,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9799965761209789,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9825029216314617,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9837949511251951,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9844128207156533,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9846622175291965,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9858586050962147,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9874829602868933,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9886003411129901,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929465667197579,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943231183447336,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9948325898302229,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9970799153180498,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9986485607529941,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.18.mlp": [
+ {
+ "accuracy": 0.8884449005126953,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8916668139005962,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9092682976471751,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9145663035543341,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9436676376744321,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9480751909707722,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9554878630136189,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9705403353038587,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9730094075202942,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9712646258504767,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9747685382240697,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.985350290411397,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9874100261612943,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9919794884167219,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9925483237755927,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940939690721663,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9978535218458426,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.19.self_attn": [
+ {
+ "accuracy": 0.9401498531040392,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9407648601030049,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9543732373337996,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9624346557416414,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9693697314513358,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9677945093104714,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9796481712868339,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9805479959437722,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9832630957427778,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.983885429407421,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9840410587034727,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9850146111689115,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9877929961995074,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9889488541766217,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930207525428972,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944936531154733,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943486757968601,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9973370654410437,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985761371882338,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.19.mlp": [
+ {
+ "accuracy": 0.8899241874092504,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.893095894863731,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9098778335671676,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9149389455192968,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9442868483693976,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9486843441662035,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9558398535377101,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9709303818250957,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9733431652972573,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.971611709971177,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.975079313704842,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9855456791426006,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9875702520734385,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9920636729190224,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9926592605678659,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940879733154648,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.997880191787293,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.20.self_attn": [
+ {
+ "accuracy": 0.9434972374062789,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9461684289731478,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9571137083204169,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9654845940439325,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9717379651571575,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9726910779350683,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9812400607686294,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9823602563456485,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9842849656155235,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9849406716070677,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9856660428800081,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9865331822320035,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9885139339848569,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9894989376005373,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937212792666335,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9948330356886512,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9951923038614424,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9973399458747161,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987123510555217,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.20.mlp": [
+ {
+ "accuracy": 0.8899101458097759,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8930120468139648,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.909466122326098,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9144694491436607,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.944171855324193,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9485053231841639,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9555640628463343,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9708365609771327,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9732858036693773,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.971556936439715,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9749964161923057,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9855262778307262,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9875438401573583,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9920980208798459,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9926450919163855,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940604979270383,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9978977745692981,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.21.self_attn": [
+ {
+ "accuracy": 0.9388996739136546,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9418948386844835,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9523384947525828,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9610453718586972,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9694268515235499,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9707936111249422,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9802439730418356,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9815980948899922,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9833786393466749,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9841321942053343,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9850632548332214,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9861579129570409,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9879073257509031,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9889393856650904,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9932869437493776,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944928582561644,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9947861393815592,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9971682417549586,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9986355621367693,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.21.mlp": [
+ {
+ "accuracy": 0.8927245265559146,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8957754185325221,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9113743367948031,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9161421938946372,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9455283942975496,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9498240320306075,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9565093611416063,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9714229357870001,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9738643749764091,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9722435474395752,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9756219857617429,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.985855199788746,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9878422264990053,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9922045720250983,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928332073123831,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9941743816200056,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9978975522282877,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.22.self_attn": [
+ {
+ "accuracy": 0.9459140488975927,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.947329979193838,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9564333213003058,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.964682328073602,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9720744867073862,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9734345486289576,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.980980490383349,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9832655411017569,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9846288357910357,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9850486030704096,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9863241252146269,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869047677830646,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9889368069799322,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9898396801007422,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937535387120748,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9948992874277266,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9951362837302057,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9973736238714895,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987216840841269,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.22.mlp": [
+ {
+ "accuracy": 0.8963537278928255,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8992149578897577,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9139569746820551,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9184379326669794,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9474278531576458,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9514634703335009,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9577942082756444,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9725929875122873,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9748857727176264,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9732493569976405,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9764631625853086,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9864166582885542,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9882749264177523,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.992601675422568,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930978558565441,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943435360726557,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980134938500429,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.23.self_attn": [
+ {
+ "accuracy": 0.9488015268978319,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9513550902667799,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9625976713080155,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9687984146569905,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.974071570132908,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9751500879463396,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9822341878163187,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9847981647441262,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.985618991287131,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9859480324544405,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9868118057125493,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9879890296020006,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9901094428802791,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9909391458097258,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944531047030499,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9954459855431005,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9954834364746746,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9977646093619498,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987865340356764,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.23.mlp": [
+ {
+ "accuracy": 0.898956763116937,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.901730543688724,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9154769433172125,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9196887706455431,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9486175399077565,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9525912529543826,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.958533945836519,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9732472771092465,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9754853781900907,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.973858925857042,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9770068604695169,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867258260124608,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9885516362754923,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9927716357143301,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9932437743011274,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944074224484595,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980438057529298,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.24.self_attn": [
+ {
+ "accuracy": 0.9468826431977122,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9502476171443337,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9601914349355196,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9672536034333079,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9731734363656295,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9743854121157998,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9826153281487917,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9833601038706931,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9847619000234102,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9852802361312666,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9862727115028783,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9872555050410723,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9894367346638128,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9902335970025313,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9941134468505257,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9951620047029696,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9952585948140997,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9974761795448629,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987592585384846,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.24.mlp": [
+ {
+ "accuracy": 0.9007913062446996,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9035158910249409,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9165905463068109,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9206223613337466,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.949508221525895,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9534020016067907,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9590853766391152,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9736765089787935,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9758791766668621,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9742876306960457,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9773814489966944,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869383204924432,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887354468044481,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928766041994095,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933525323867798,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944581844304737,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980714477990803,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.25.self_attn": [
+ {
+ "accuracy": 0.9467459979810213,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9497765835962797,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9588482630880255,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9654307114450555,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9723404990999323,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9735176029958224,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9822318177474173,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9834140021550027,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9849763481240523,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9855139679030368,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9862264865323117,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9875599687036715,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9894656295839109,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9904759487039164,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9941226236130062,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9951061155450972,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9952703255571818,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9974095411598682,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987451795880732,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.25.mlp": [
+ {
+ "accuracy": 0.9012252594295301,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9039818801377949,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.91665560320804,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9205681273811742,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9497011586239463,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.953595685331445,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9590949949465299,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9737410482607389,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9759437426140434,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9743860363960266,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9774677031918576,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986985729713189,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887751333023372,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.992900793489657,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933874379647406,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944590565405393,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980818578287175,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.26.self_attn": [
+ {
+ "accuracy": 0.9466159469202945,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.948185174088729,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9571042374560708,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9652439857784071,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.97223751168502,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9735791244004902,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9824786954804471,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9834249552927519,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9847503618190163,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9854327863768527,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9860591606089943,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869148715546257,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9889599045640544,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9899848644670687,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9938989442430044,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9949883857839986,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9952832503538382,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.997303755855874,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9987262575642059,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.26.mlp": [
+ {
+ "accuracy": 0.9008493925395765,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9035610525231612,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9161956561239142,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9201457876908152,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9494300516028154,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9533560025064569,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9588515915368733,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9735928271946154,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9757954682174482,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9742584620651445,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9773390653886294,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869172886798256,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887142314722663,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9928427777792278,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933555581067738,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9944306408104143,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980642487736124,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.27.self_attn": [
+ {
+ "accuracy": 0.9408501763092845,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9408272065614399,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9541349630606801,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9623476737423947,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.970335747066297,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.971617413194556,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9805558151320407,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9818708065309023,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9829453559298265,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9837872825170818,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9851985228689093,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9858796439672771,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9880004634982661,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9890461846401817,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933841942172301,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9945340238903698,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9947667655191923,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9970771944836566,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985992286942507,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.27.mlp": [
+ {
+ "accuracy": 0.8994839818854081,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9023114819275706,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.914894373793351,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.918943769053409,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9486716044576544,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9526794897882562,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9582008186139559,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9730287539331537,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9753501211342058,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9738497906609586,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9770147863187288,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9867166986590937,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9885519742965698,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9926954829379132,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9932587797704496,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943423965259602,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9980315417051315,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.28.self_attn": [
+ {
+ "accuracy": 0.9319704457333213,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9351067229321128,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9487236549979762,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9597707955460799,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9663907195392408,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9665976794142472,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9797934685882769,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.980740958138516,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9821813310447493,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.983206171738474,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9830496028849953,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9841822385787964,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9869235871653808,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9880931149972113,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.992810416378473,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9939277740685564,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.994602664520866,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9969771084816832,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9985299534899624,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.28.mlp": [
+ {
+ "accuracy": 0.8952982174722772,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8982489485489695,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9112675440938849,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9155886863407335,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9465327764812269,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9506916403770447,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.956439799384067,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9718038784830194,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9742104050360227,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9727693105998793,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9760473841115048,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9861613684578946,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9880636718712354,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9923684353891172,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9929744154214859,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9941188112685555,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9979268922225425,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.29.self_attn": [
+ {
+ "accuracy": 0.935006060098347,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9363104104995728,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.951394564227054,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9595166601632771,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9668202839399639,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9678617904060766,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9800648548101124,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9806999438687375,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9829893018070021,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9835452625626012,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9836794247752741,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9844651457510496,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.987778514623642,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9887642146725404,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9933119392708728,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9943140745162964,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9949515305067363,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9970791820240649,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9986073457096752,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.29.mlp": [
+ {
+ "accuracy": 0.8888981718766062,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8920954277640895,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.905916414762798,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9105437994003296,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.942714910758169,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9475349595672206,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9536912629478856,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9696492019452547,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9723044508381894,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9705999368115475,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9743707917238537,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9849619238000167,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9871856350647775,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9915788236417269,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9923689671252903,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9935894294788963,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.997731566821274,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.30.self_attn": [
+ {
+ "accuracy": 0.9243166509427523,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9271602818840429,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9416637295170834,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9521459905724776,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9627030209491128,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9631254610262419,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9771877433124342,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9781057097409901,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9798939824104309,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9809304664009496,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9815213099906319,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9826202172982065,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9856778540109333,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.986888491009411,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.992218637152722,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9934052503422687,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9940707903159293,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9965660552445211,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9984502180626518,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.30.mlp": [
+ {
+ "accuracy": 0.8761413348348517,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.880308515147159,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8944236228340551,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8991182226883738,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9350596603594328,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9417252289621454,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9480985904994764,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9661526554509213,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.969013860351161,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9664685036006727,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9714578766571847,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9828246126049444,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9857153798404493,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9903610215375298,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9912104575257552,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9924603873177579,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9973659205593561,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.31.self_attn": [
+ {
+ "accuracy": 0.921551177376195,
+ "total_bits": 89665536,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9259662000756514,
+ "total_bits": 92221440,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9415510955609774,
+ "total_bits": 95758848,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9563444131299069,
+ "total_bits": 112272384,
+ "q_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9620950033790187,
+ "total_bits": 132913152,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9623824232502988,
+ "total_bits": 132980224,
+ "q_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64,
+ "3": 64
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9756921121948644,
+ "total_bits": 169613312,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9767073016417654,
+ "total_bits": 169745920,
+ "q_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.978857590963966,
+ "total_bits": 171195392,
+ "q_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 64
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.97982257604599,
+ "total_bits": 173563904,
+ "q_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9810058568653307,
+ "total_bits": 174923264,
+ "q_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9821856178735432,
+ "total_bits": 175750144,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9847583927606282,
+ "total_bits": 179253248,
+ "q_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 64
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 64,
+ "4": 64
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9862148134331954,
+ "total_bits": 181592064,
+ "q_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "5": 32
+ },
+ "bits": [
+ 5
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9917345980280324,
+ "total_bits": 220469248,
+ "q_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9930143434750406,
+ "total_bits": 223535104,
+ "q_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9937065004518157,
+ "total_bits": 253499392,
+ "q_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9962748932211023,
+ "total_bits": 265838592,
+ "q_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 32
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "6": 32
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9983115135446975,
+ "total_bits": 337385472,
+ "q_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "k_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "v_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "o_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.layers.31.mlp": [
+ {
+ "accuracy": 0.8429100262491327,
+ "total_bits": 395461696,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8491878760488409,
+ "total_bits": 409224256,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "6": 32,
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 6,
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.2,
+ 0.75
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8640761124460321,
+ "total_bits": 457272320,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "3": 32
+ },
+ "bits": [
+ 5,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.8689453099903307,
+ "total_bits": 512977920,
+ "gate_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "3": 64,
+ "2": 64
+ },
+ "bits": [
+ 3,
+ 2
+ ],
+ "bits_prop": [
+ 0.3,
+ 0.7
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9211063259526303,
+ "total_bits": 578397280,
+ "gate_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128,
+ "3": 128
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9291951405374628,
+ "total_bits": 592872448,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 8,
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9356453167764764,
+ "total_bits": 637454432,
+ "gate_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32,
+ "3": 32
+ },
+ "bits": [
+ 4,
+ 3
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9586272992585835,
+ "total_bits": 728741472,
+ "gate_proj": {
+ "group_size": {
+ "4": 128
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9622115084999486,
+ "total_bits": 739476480,
+ "gate_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "4": 32
+ },
+ "bits": [
+ 4
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9591140151023865,
+ "total_bits": 751543392,
+ "gate_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 128,
+ "4": 128
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9651411363953039,
+ "total_bits": 766018560,
+ "gate_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "5": 32,
+ "4": 32
+ },
+ "bits": [
+ 8,
+ 5,
+ 4
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9787399988425406,
+ "total_bits": 924689504,
+ "gate_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128,
+ "5": 128
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9823066388305864,
+ "total_bits": 939164672,
+ "gate_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.25,
+ 0.75
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 32,
+ "5": 32
+ },
+ "bits": [
+ 8,
+ 6,
+ 5
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.1,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9878395748765845,
+ "total_bits": 1069524064,
+ "gate_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "6": 128
+ },
+ "bits": [
+ 6
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 32,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.05,
+ 0.95
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.988988533616066,
+ "total_bits": 1110384896,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.15,
+ 0.85
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9903163462877274,
+ "total_bits": 1209999616,
+ "gate_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128,
+ "6": 128
+ },
+ "bits": [
+ 8,
+ 6
+ ],
+ "bits_prop": [
+ 0.1,
+ 0.9
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ },
+ {
+ "accuracy": 0.9962382257768982,
+ "total_bits": 1415520512,
+ "gate_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "up_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ },
+ "down_proj": {
+ "group_size": {
+ "8": 128
+ },
+ "bits": [
+ 8
+ ],
+ "bits_prop": [
+ 1
+ ],
+ "scale_bits": 4
+ }
+ }
+ ],
+ "model.norm.norm": null,
+ "lm_head.linear": null
+ },
+ "last_module_idx": 66
+}
\ No newline at end of file