File size: 2,999 Bytes
e006462 daf9278 658687b e3d0951 148c4b9 309501d 148c4b9 309501d 148c4b9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 |
---
base_model:
- unsloth/Mistral-Nemo-Instruct-2407
---
Note: This model is no longer the optimal W8A8 quantization, please consider using a better quantization model I made later:
noneUsername/Mistral-Nemo-Instruct-2407-W8A8-Dynamic-Per-Token-better
My first quantization uses the quantization method provided by vllm:
https://docs.vllm.ai/en/latest/quantization/int8.html
NUM_CALIBRATION_SAMPLES = 2048
MAX_SEQUENCE_LENGTH = 8192
smoothing_strength=0.8
I will verify the validity of the model and update the readme as soon as possible.
edit: The performance in my ERP test was comparable to Mistral-Nemo-Instruct-2407-GPTQ-INT8, which I consider a successful quantization.
vllm (pretrained=/root/autodl-tmp/Mistral-Nemo-Instruct-2407,add_bos_token=true,tensor_parallel_size=2,max_model_len=4096,gpu_memory_utilization=0.85,swap_space=0), gen_kwargs: (None), limit: 250.0, num_fewshot: 5, batch_size: auto
|Tasks|Version| Filter |n-shot| Metric | |Value| |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.800|± |0.0253|
| | |strict-match | 5|exact_match|↑ |0.784|± |0.0261|
lm_eval --model vllm --model_args pretrained="/mnt/e/Code/models/Mistral-Nemo-Instruct-2407-W8A8-Dynamic-Per-Token",add_bos_token=true,dtype=half,tensor_parallel_size=2,max_model_len=4096,gpu_memory_utilization=0.85,swap_space=0 --tasks gsm8k --num_fewshot 5 --limit 250 --batch_size 1
vllm (pretrained=/mnt/e/Code/models/Mistral-Nemo-Instruct-2407-W8A8-Dynamic-Per-Token,add_bos_token=true,dtype=half,tensor_parallel_size=2,max_model_len=4096,gpu_memory_utilization=0.85,swap_space=0), gen_kwargs: (None), limit: 250.0, num_fewshot: 5, batch_size: 1
|Tasks|Version| Filter |n-shot| Metric | |Value| |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.784|± |0.0261|
| | |strict-match | 5|exact_match|↑ |0.768|± |0.0268|
In gsm8k, still a bit worse than the original...
lm_eval --model vllm \
l_args > --model_args pretrained="/mnt/e/Code/models/Mistral-Nemo-Instruct-2407-W8A8-Dynamic-Per-Token",add_bos_token=true,dtype=half,tensor_parallel_size=2,max_model_len=4096,gpu_memory_utilization=0.85,swap_space=0 \
ks hellaswag \
> --tasks hellaswag \
> --limit 150 \
> --num_fewshot 10 \
--batch_size 1
vllm (pretrained=/mnt/e/Code/models/Mistral-Nemo-Instruct-2407-W8A8-Dynamic-Per-Token,add_bos_token=true,dtype=half,tensor_parallel_size=2,max_model_len=4096,gpu_memory_utilization=0.85,swap_space=0), gen_kwargs: (None), limit: 150.0, num_fewshot: 10, batch_size: 1
| Tasks |Version|Filter|n-shot| Metric | |Value | |Stderr|
|---------|------:|------|-----:|--------|---|-----:|---|-----:|
|hellaswag| 1|none | 10|acc |↑ |0.5800|± |0.0404|
| | |none | 10|acc_norm|↑ |0.7533|± |0.0353| |