cognitivecomputations

/

DeepSeek-R1-AWQ

like 54

Follow

Cognitive Computations 1.61k

Hello, I am using vllm to deploy the inference service. The usage.prompt_tokens data returned by the calling interface is inconsistent with the token obtained using transformers.AutoTokenizer. The following is the test process:
vllm startup command:

docker run -itd \
  --name deepseek-awq \
  --network host \
  --shm-size=1024m \
  --gpus all \
  -v $(pwd):/app \
  --entrypoint "bash" \
  docker.1ms.run/vllm/vllm-openai:v0.7.2 \
  -c " python3 -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000 --enable_prefix_caching  --max-model-len 65536 --trust-remote-code --tensor-parallel-size 8 --quantization moe_wna16 --gpu-memory-util 0.97 --kv-cache-dtype fp8_e5m2 --calculate-kv-scales --served-model-name deepseek-awq --enable-chunked-prefill --model /app/cognitivecomputations/DeepSeek-R1-awq"

curl --request POST \
  --url http://10.1.30.59:8000/v1/chat/completions \
  --header 'Authorization: Bearer sk-ddddddddddddddd' \
  --header 'Content-Type: application/json' \
  --data '{
    "messages": [
        {
            "role": "user",
            "content": "what is you name"
        }
    ],
    "stream": false,
    // "stream_options":{
    //     "include_usage": true
    // },
    "model": "deepseek-awq",
    "temperature": 0.5,
    "presence_penalty": 0,
    "frequency_penalty": 0,
    "top_p": 1
}'


Response:
    "usage": {
        "prompt_tokens": 7,
        "total_tokens": 120,
        "completion_tokens": 113,
        "prompt_tokens_details": null
    },

python code:

# pip3 install transformers
# python3 deepseek_tokenizer.py
import transformers

chat_tokenizer_dir = "./"

tokenizer = transformers.AutoTokenizer.from_pretrained( 
        chat_tokenizer_dir, trust_remote_code=True
        )

result = tokenizer.encode("what is you name")
print(result)

root@A100-GPU-59:/app/cognitivecomputations/DeepSeek-R1-awq# python3 deepseek_tokenizer.py 
[0, 9602, 344, 440, 2329]

They use the same tokenizer configuration