Update README.md
#6
by
haoyang-amd
- opened
README.md
CHANGED
|
@@ -8,7 +8,7 @@ base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
|
|
| 8 |
- ## Introduction
|
| 9 |
This model was created by applying [Quark](https://quark.docs.amd.com/latest/index.html) with calibration samples from Pile dataset.
|
| 10 |
- ## Quantization Stragegy
|
| 11 |
-
- ***Quantized Layers***: All linear layers excluding "lm_head", "
|
| 12 |
- ***Weight***: FP8 symmetric per-tensor
|
| 13 |
- ***Activation***: FP8 symmetric per-tensor
|
| 14 |
- ***KV Cache***: FP8 symmetric per-tensor
|
|
@@ -23,8 +23,9 @@ python3 quantize_quark.py \
|
|
| 23 |
--output_dir Mixtral-8x7B-Instruct-v0.1-FP8-KV \
|
| 24 |
--quant_scheme w_fp8_a_fp8 \
|
| 25 |
--kv_cache_dtype fp8 \
|
| 26 |
-
--num_calib_data 128
|
| 27 |
-
--model_export quark_safetensors
|
|
|
|
| 28 |
# If model size is too large for single GPU, please use multi GPU instead.
|
| 29 |
python3 quantize_quark.py \
|
| 30 |
--model_dir $MODEL_DIR \
|
|
@@ -33,6 +34,7 @@ python3 quantize_quark.py \
|
|
| 33 |
--kv_cache_dtype fp8 \
|
| 34 |
--num_calib_data 128 \
|
| 35 |
--model_export quark_safetensors \
|
|
|
|
| 36 |
--multi_gpu
|
| 37 |
```
|
| 38 |
## Deployment
|
|
@@ -53,9 +55,9 @@ The quantization evaluation results are conducted in pseudo-quantization mode, w
|
|
| 53 |
<tr>
|
| 54 |
<td>Perplexity-wikitext2
|
| 55 |
</td>
|
| 56 |
-
<td>4.
|
| 57 |
</td>
|
| 58 |
-
<td>4.
|
| 59 |
</td>
|
| 60 |
</tr>
|
| 61 |
</table>
|
|
|
|
| 8 |
- ## Introduction
|
| 9 |
This model was created by applying [Quark](https://quark.docs.amd.com/latest/index.html) with calibration samples from Pile dataset.
|
| 10 |
- ## Quantization Stragegy
|
| 11 |
+
- ***Quantized Layers***: All linear layers excluding "lm_head", "*.gate"
|
| 12 |
- ***Weight***: FP8 symmetric per-tensor
|
| 13 |
- ***Activation***: FP8 symmetric per-tensor
|
| 14 |
- ***KV Cache***: FP8 symmetric per-tensor
|
|
|
|
| 23 |
--output_dir Mixtral-8x7B-Instruct-v0.1-FP8-KV \
|
| 24 |
--quant_scheme w_fp8_a_fp8 \
|
| 25 |
--kv_cache_dtype fp8 \
|
| 26 |
+
--num_calib_data 128 \
|
| 27 |
+
--model_export quark_safetensors \
|
| 28 |
+
--no_weight_matrix_merge
|
| 29 |
# If model size is too large for single GPU, please use multi GPU instead.
|
| 30 |
python3 quantize_quark.py \
|
| 31 |
--model_dir $MODEL_DIR \
|
|
|
|
| 34 |
--kv_cache_dtype fp8 \
|
| 35 |
--num_calib_data 128 \
|
| 36 |
--model_export quark_safetensors \
|
| 37 |
+
--no_weight_matrix_merge \
|
| 38 |
--multi_gpu
|
| 39 |
```
|
| 40 |
## Deployment
|
|
|
|
| 55 |
<tr>
|
| 56 |
<td>Perplexity-wikitext2
|
| 57 |
</td>
|
| 58 |
+
<td>4.1387
|
| 59 |
</td>
|
| 60 |
+
<td>4.2207
|
| 61 |
</td>
|
| 62 |
</tr>
|
| 63 |
</table>
|