Upload folder using huggingface_hub

16_03_i2v_140_tgst1k_loras/epoch20/16_03_i2v_140_tgst1k_config.toml

@@ -0,0 +1,60 @@
+output_dir = '/workspace/output_models/16_03_i2v_140_tgst1k_model'
+
+# Dataset config file.
+dataset = 'examples/16_03_i2v_140_tgst1k_dataset.toml'
+
+# training settings
+
+epochs = 30
+micro_batch_size_per_gpu = 1
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+
+# misc settings
+
+save_every_n_epochs = 5
+checkpoint_every_n_minutes = 20
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+save_dtype = 'bfloat16'
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 1
+steps_per_print = 1
+video_clip_mode = 'single_beginning'
+
+[model]
+type = 'wan'
+ckpt_path = '/workspace/models/wan21i2v'
+dtype = 'bfloat16'
+timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'adamw_optimi'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

16_03_i2v_140_tgst1k_loras/epoch20/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v",
+    "ffn.0",
+    "k_img",
+    "ffn.2",
+    "o",
+    "q",
+    "v_img",
+    "k"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

16_03_i2v_140_tgst1k_loras/epoch20/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50f314040fb9610c7bf33696fb205bdb7006c438f20fe34afd9bca75a3d421a5
+size 359257680

16_03_i2v_140_tgst1k_loras/epoch25/16_03_i2v_140_tgst1k_config.toml

@@ -0,0 +1,60 @@
+output_dir = '/workspace/output_models/16_03_i2v_140_tgst1k_model'
+
+# Dataset config file.
+dataset = 'examples/16_03_i2v_140_tgst1k_dataset.toml'
+
+# training settings
+
+epochs = 30
+micro_batch_size_per_gpu = 1
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+
+# misc settings
+
+save_every_n_epochs = 5
+checkpoint_every_n_minutes = 20
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+save_dtype = 'bfloat16'
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 1
+steps_per_print = 1
+video_clip_mode = 'single_beginning'
+
+[model]
+type = 'wan'
+ckpt_path = '/workspace/models/wan21i2v'
+dtype = 'bfloat16'
+timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'adamw_optimi'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

16_03_i2v_140_tgst1k_loras/epoch25/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v",
+    "ffn.0",
+    "k_img",
+    "ffn.2",
+    "o",
+    "q",
+    "v_img",
+    "k"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

16_03_i2v_140_tgst1k_loras/epoch25/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d64467f293333c7540ebc50ba88d3cf6e69de96732deb645f68beda0a5a0d1b
+size 359257680

16_03_i2v_140_tgst1k_loras/epoch30/16_03_i2v_140_tgst1k_config.toml

@@ -0,0 +1,60 @@
+output_dir = '/workspace/output_models/16_03_i2v_140_tgst1k_model'
+
+# Dataset config file.
+dataset = 'examples/16_03_i2v_140_tgst1k_dataset.toml'
+
+# training settings
+
+epochs = 30
+micro_batch_size_per_gpu = 1
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+
+# misc settings
+
+save_every_n_epochs = 5
+checkpoint_every_n_minutes = 20
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+save_dtype = 'bfloat16'
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 1
+steps_per_print = 1
+video_clip_mode = 'single_beginning'
+
+[model]
+type = 'wan'
+ckpt_path = '/workspace/models/wan21i2v'
+dtype = 'bfloat16'
+timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'adamw_optimi'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

16_03_i2v_140_tgst1k_loras/epoch30/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v",
+    "ffn.0",
+    "k_img",
+    "ffn.2",
+    "o",
+    "q",
+    "v_img",
+    "k"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

16_03_i2v_140_tgst1k_loras/epoch30/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e772dfc14bd645efc151aac397968755bbe320e201b829752aacf3f3cbd4d406
+size 359257680

16_03_i2v_140_tgst1k_loras/trigger.txt

@@ -1 +0,0 @@
-tgst1k style