Commit
•
52efb3a
1
Parent(s):
62f74b5
Update README.md (#9)
Browse files- Update README.md (d0c8391aeee79a2c97670ecd97b5d35df548412d)
Co-authored-by: haoyanli <haoyang-amd@users.noreply.huggingface.co>
README.md
CHANGED
|
@@ -14,7 +14,7 @@ base_model:
|
|
| 14 |
- ## Introduction
|
| 15 |
This model was created by applying [Quark](https://quark.docs.amd.com/latest/index.html) with calibration samples from Pile dataset.
|
| 16 |
- ## Quantization Stragegy
|
| 17 |
-
- ***Quantized Layers***: All linear layers excluding "lm_head", "
|
| 18 |
- ***Weight***: FP8 symmetric per-tensor
|
| 19 |
- ***Activation***: FP8 symmetric per-tensor
|
| 20 |
- ***KV Cache***: FP8 symmetric per-tensor
|
|
@@ -29,16 +29,18 @@ python3 quantize_quark.py \
|
|
| 29 |
--output_dir deepseek-moe-16b-chat-FP8-KV \
|
| 30 |
--quant_scheme w_fp8_a_fp8 \
|
| 31 |
--kv_cache_dtype fp8 \
|
| 32 |
-
--num_calib_data 128
|
| 33 |
-
--model_export quark_safetensors
|
|
|
|
| 34 |
# If model size is too large for single GPU, please use multi GPU instead.
|
| 35 |
python3 quantize_quark.py \
|
| 36 |
--model_dir $MODEL_DIR \
|
| 37 |
--output_dir deepseek-moe-16b-chat-FP8-KV \
|
| 38 |
--quant_scheme w_fp8_a_fp8 \
|
| 39 |
--kv_cache_dtype fp8 \
|
| 40 |
-
--num_calib_data 128
|
| 41 |
--model_export quark_safetensors \
|
|
|
|
| 42 |
--multi_gpu
|
| 43 |
```
|
| 44 |
## Deployment
|
|
|
|
| 14 |
- ## Introduction
|
| 15 |
This model was created by applying [Quark](https://quark.docs.amd.com/latest/index.html) with calibration samples from Pile dataset.
|
| 16 |
- ## Quantization Stragegy
|
| 17 |
+
- ***Quantized Layers***: All linear layers excluding "lm_head", "*.gate"
|
| 18 |
- ***Weight***: FP8 symmetric per-tensor
|
| 19 |
- ***Activation***: FP8 symmetric per-tensor
|
| 20 |
- ***KV Cache***: FP8 symmetric per-tensor
|
|
|
|
| 29 |
--output_dir deepseek-moe-16b-chat-FP8-KV \
|
| 30 |
--quant_scheme w_fp8_a_fp8 \
|
| 31 |
--kv_cache_dtype fp8 \
|
| 32 |
+
--num_calib_data 128 \
|
| 33 |
+
--model_export quark_safetensors \
|
| 34 |
+
--no_weight_matrix_merge
|
| 35 |
# If model size is too large for single GPU, please use multi GPU instead.
|
| 36 |
python3 quantize_quark.py \
|
| 37 |
--model_dir $MODEL_DIR \
|
| 38 |
--output_dir deepseek-moe-16b-chat-FP8-KV \
|
| 39 |
--quant_scheme w_fp8_a_fp8 \
|
| 40 |
--kv_cache_dtype fp8 \
|
| 41 |
+
--num_calib_data 128 \
|
| 42 |
--model_export quark_safetensors \
|
| 43 |
+
--no_weight_matrix_merge \
|
| 44 |
--multi_gpu
|
| 45 |
```
|
| 46 |
## Deployment
|