deepseek.md

DeepSeek

DeepSeek-R1-Distill Models

The distilled models can be used in the same way as the base model. For instance:

• DeepSeek-R1-Distill-Qwen-xxB, based on Qwen2.5, utilizes Qwen2Convert.
• DeepSeek-R1-Distill-Llama-xxB employs LlamaConvert.

DeepSeek-R1 671B

Notice: DeepSeek-R1 671B only supports dtype fp8_e4m3 and kvcache dtype bf16.

Requirements

• xfastertransformer >= 2.0.0
• oneCCL
• vllm-xft >= 0.5.5.3

Running Benchmarks

Execute the benchmark scripts to evaluate the model's performance using fake weight without downloading 600GB+ weight.

    git clone https://github.com/intel/xFasterTransformer.git
    cd xFasterTransformer/benchmark

    export $(python -c 'import xfastertransformer as xft; print(xft.get_env())')
    bash run_benchmark.sh -m deepseek-r1 -d fp8_e4m3 -kvd bf16 -bs 1 -in 32 -out 32 -s 1

• -bs: batch size.
• -in 32: input token length, [32, 64, 128, 256, 512, 1024, 2048, 4096, 8192].
• -out: output length.
• -s: number of used CPU nodes. If you only have 1 node with SNC-3 enabled, it will be used as number of subnumas.

Serving the Model(OpenAI Compatible Server)

Preparing the Model

• Download original DeepSeek-R1 671B model from HuggingFace.
• Convert model into xFT format.

  python -c 'import xfastertransformer as xft; xft.DeepSeekR1Convert().convert("${HF_DATASET_DIR}","${OUTPUT_DIR}")'
  

  After conversion, the *.safetensor files in ${HF_DATASET_DIR} is no longer needed if you want to save storage space.

Run Serving

• Single instance
  If you want to run DeepSeek in one CPU numa, like NUMA node0 CPU(s): 0-47,96-143

  # Preload libiomp5.so by following cmd or LD_PRELOAD=libiomp5.so manually
  export $(python -c 'import xfastertransformer as xft; print(xft.get_env())')
  
  numactl 0-47 -l python -m vllm.entrypoints.openai.api_server \
          --model ${MODEL_PATH} \
          --tokenizer ${TOKEN_PATH} \
          --dtype fp8_e4m3 \
          --kv-cache-dtype bf16 \
          --served-model-name xft \
          --port 8000 \
          --trust-remote-code 

  • MODEL_PATH: The xFT format model weights.
  • TOKEN_PATH: The tokenizer related files, like HF_DATASET_DIR.
  • served-model-name: The model name used in the API.

• Distributed(Multi-rank) If you want to run DeepSeek cross numas, like

  NUMA node0 CPU(s):   0-47,96-143
  NUMA node1 CPU(s):   48-95,144-191
  

  # Preload libiomp5.so by following cmd or LD_PRELOAD=libiomp5.so manually
  export $(python -c 'import xfastertransformer as xft; print(xft.get_env())')
  
  OMP_NUM_THREADS=48 mpirun \
          -n 1 numactl --all -C 0-47 \
          python -m vllm.entrypoints.openai.api_server \
              --model ${MODEL_PATH} \
              --tokenizer ${TOKEN_PATH} \
              --dtype fp8_e4m3 \
              --kv-cache-dtype bf16 \
              --served-model-name xft \
              --port 8000 \
              --trust-remote-code \
          : -n 1 numactl --all -C 48-95 \
          python -m vllm.entrypoints.slave \
              --dtype fp8_e4m3 \
              --model ${MODEL_PATH} \
              --kv-cache-dtype bf16

• Query example

  curl http://localhost:8000/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
          "model": "xft",
          "messages": [{"role": "user", "content": "你好呀！请问你是谁？"}],
          "max_tokens": 256,
          "temperature": 0.6,
          "top_p": 0.95
      }'