I don’t use Docker and want to deploy deepseek-v4 locally using pip, but the latest version of vllm is 0.20. Do I need to wait for version 0.20.1 to be released? The GitHub source code compilation shows that the transformer is outdated.
You do not need to wait for vLLM 0.20.1. To run DeepSeek-V4 locally, you should install the latest vLLM nightly build (not the stable 0.20) and also install the latest version of transformers from source, as DeepSeek-V4 requires features and fixes not present in the current stable releases. The error about transformers being outdated is expected if you use the default pip version; you must upgrade transformers manually after installing vLLM nightly.
Use these commands:
uv pip install -U vllm --extra-index-url https://wheels.vllm.ai/nightly
pip install git+https://github.com/huggingface/transformers.git
This approach is confirmed in the official recipes and GitHub issues for DeepSeek-V4 and recent models, where users are instructed to use nightly wheels and the latest transformers from source for compatibility and bug fixes. Would you like more detail on troubleshooting or configuration?
Sources:
Hint: Mention @RunLLM in the post for followups.
Note, DeepSeek V4 is supported in v0.20. If there is a specific issue fixed on main since then, you can use nightly as suggested above