yusufs/vllm-inference

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

poetry export -f requirements.txt --output requirements.txt --without-hashes

• —The HUGGING_FACE_HUB_TOKEN and HF_TOKEN must exist during runtime (use the same value, it must have read permission to the model.)

VLLM OpenAI Compatible API Server

References: https://huggingface.co/spaces/sofianhw/ai/tree/c6527a750644a849b6705bb6fe2fcea4e54a8196

This api_server.py file is exact copy version from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/entrypoints/openai/api_server.py

Changes (use diff tool to see the exact changes of the file):

• —[x] change everything route in apiserver.py that start (“/v1/xxx”) to (“/api/v1/xxx”). and just run the python apiserver.py with arguments. https://discuss.huggingface.co/t/run-vllm-docker-on-space/70228/5?u=yusufs

Documentation about config

• —https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/utils.py#L1207-L1221

"serve,chat,complete",
"facebook/opt-12B",
'--config', 'config.yaml',
'-tp', '2'

The yaml is equivalent with argument flag params. Consider passing using flag params that defined here for better documentation: https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/entrypoints/openai/cli_args.py#L77-L237

Other arguments is the same as LLM class such as --max-model-len, --dtype, or --otlp-traces-endpoint

• —https://github.com/vllm-project/vllm/blob/v0.6.4/vllm/config.py#L1061-L1086
• —https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/engine/arg_utils.py#L221-L913