キサラギの世界

1
conda create -p /data/<username>/envs/qwen python=3.10 -y
2
conda activate /data/<username>/envs/qwen

1
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 \
2
  --index-url https://download.pytorch.org/whl/cu121
3
pip install vllm==0.6.6.post1 \
4
  -i https://pypi.tuna.tsinghua.edu.cn/simple
5
pip install -U huggingface_hub \
6
  -i https://pypi.tuna.tsinghua.edu.cn/simple

1
pip uninstall -y transformers tokenizers
2
pip install transformers==4.45.2 tokenizers==0.20.3 \
3
-i https://pypi.tuna.tsinghua.edu.cn/simple

1
export HF_ENDPOINT=https://hf-mirror.com
2
hf download Qwen/Qwen2.5-7B-Instruct \
3
  --local-dir /data/<username>/models/Qwen2.5-7B-Instruct

1
export HF_HUB_OFFLINE=1
2
export TRANSFORMERS_OFFLINE=1

1
CUDA_VISIBLE_DEVICES=0,1 vllm serve \
2
  /data/<username>/models/Qwen2.5-7B-Instruct \
3
  --host 0.0.0.0 \
4
  --port 8002 \
5
  --served-model-name chat-model \
6
  --tensor-parallel-size 2

1
curl http://127.0.0.1:8002/v1/models
2
curl http://127.0.0.1:8002/v1/chat/completions \
3
  -H "Content-Type: application/json" \
4
  -d '{
5
    "model":"chat-model",
6
    "messages":[
7
      {
8
        "role":"user",
9
        "content":"介绍一下后训练中的GRPO算法"
10
      }
11
    ]
12
  }'