| license: mit | |
| language: | |
| - en | |
| - zh | |
| base_model: | |
| - deepseek-ai/DeepSeek-R1 | |
| pipeline_tag: text-generation | |
| library_name: transformers | |
| # DeepSeek R1 AWQ | |
| AWQ of DeepSeek R1. | |
| This quant modified some of the model code to fix an overflow issue when using float16. | |
| To serve using vLLM with 8x 80GB GPUs, use the following command: | |
| ```sh | |
| python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 12345 --max-model-len 65536 --trust-remote-code --tensor-parallel-size 8 --quantization moe_wna16 --gpu-memory-utilization 0.97 --kv-cache-dtype fp8_e5m2 --calculate-kv-scales --served-model-name deepseek-reasoner --model cognitivecomputations/DeepSeek-R1-AWQ | |
| ``` | |
| The max model length flag ensures that KV cache usage won't be higher than available memory, the `moe_wna16` kernel doubles the inference speed, but you must build vLLM from source as of 2025/2/3. \ | |
| You can download the wheel I built for PyTorch 2.6, Python 3.12 by clicking [here](https://huggingface.co/x2ray/wheels/resolve/main/vllm-0.7.1.dev69%2Bg4f4d427a.d20220101.cu126-cp312-cp312-linux_x86_64.whl). | |
| Inference speed with batch size 1 and short prompt: | |
| - 8x H100: 34 TPS | |
| - 8x A100: 27 TPS |