Update README.md
Browse files
README.md
CHANGED
|
@@ -61,4 +61,34 @@ datasets:
|
|
| 61 |
<td><strong>35.60</strong></td>
|
| 62 |
<td>69.43</td>
|
| 63 |
</tr>
|
| 64 |
-
</table>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
<td><strong>35.60</strong></td>
|
| 62 |
<td>69.43</td>
|
| 63 |
</tr>
|
| 64 |
+
</table>
|
| 65 |
+
|
| 66 |
+
# Quant. preference
|
| 67 |
+
|
| 68 |
+
| Quant. | Description |
|
| 69 |
+
|---------------|---------------------------------------------------------------------------------------|
|
| 70 |
+
| not_quantized | Recommended. Fast conversion. Slow inference, big files. |
|
| 71 |
+
| fast_quantized| Recommended. Fast conversion. OK inference, OK file size. |
|
| 72 |
+
| quantized | Recommended. Slow conversion. Fast inference, small files. |
|
| 73 |
+
| f32 | Not recommended. Retains 100% accuracy, but super slow and memory hungry. |
|
| 74 |
+
| f16 | Fastest conversion + retains 100% accuracy. Slow and memory hungry. |
|
| 75 |
+
| q8_0 | Fast conversion. High resource use, but generally acceptable. |
|
| 76 |
+
| q4_k_m | Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K |
|
| 77 |
+
| q5_k_m | Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K |
|
| 78 |
+
| q2_k | Uses Q4_K for the attention.vw and feed_forward.w2 tensors, Q2_K for the other tensors.|
|
| 79 |
+
| q3_k_l | Uses Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K |
|
| 80 |
+
| q3_k_m | Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K |
|
| 81 |
+
| q3_k_s | Uses Q3_K for all tensors |
|
| 82 |
+
| q4_0 | Original quant method, 4-bit. |
|
| 83 |
+
| q4_1 | Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.|
|
| 84 |
+
| q4_k_s | Uses Q4_K for all tensors |
|
| 85 |
+
| q4_k | alias for q4_k_m |
|
| 86 |
+
| q5_k | alias for q5_k_m |
|
| 87 |
+
| q5_0 | Higher accuracy, higher resource usage and slower inference. |
|
| 88 |
+
| q5_1 | Even higher accuracy, resource usage and slower inference. |
|
| 89 |
+
| q5_k_s | Uses Q5_K for all tensors |
|
| 90 |
+
| q6_k | Uses Q8_K for all tensors |
|
| 91 |
+
| iq2_xxs | 2.06 bpw quantization |
|
| 92 |
+
| iq2_xs | 2.31 bpw quantization |
|
| 93 |
+
| iq3_xxs | 3.06 bpw quantization |
|
| 94 |
+
| q3_k_xs | 3-bit extra small quantization |
|