vllm / CLAUDE.md
davanstrien's picture
davanstrien HF Staff
Sync from GitHub via hub-sync
8cba5fe verified

vLLM Scripts Development Notes

Repository Purpose

This repository contains UV scripts for vLLM-based inference tasks. Focus on GPU-accelerated inference using vLLM's optimized engine.

Key Patterns

1. GPU Requirements

All scripts MUST check for GPU availability:

if not torch.cuda.is_available():
    logger.error("CUDA is not available. This script requires a GPU.")
    sys.exit(1)

2. vLLM Docker Image

Run on the pinned vllm/vllm-openai:0.22.1 image for HF Jobs (ships vLLM + the CUDA toolkit, so FlashInfer works) with --python /usr/bin/python3 -e PYTHONPATH=/usr/local/lib/python3.12/dist-packages. Pin the tag (test-then-pin), not :latest. Tested 2026-06-05 on l4x1 (vLLM 0.22.1).

The custom nightly / FlashInfer [[tool.uv.index]] blocks (§3) are no longer needed — the image provides vLLM; keep a vllm>=0.11.0 floor in the PEP 723 header for the local path.

APIs (verified): structured outputs use StructuredOutputsParams (structured_outputs=, ≥0.11.0 — GuidedDecodingParams/guided_decoding= removed in 0.12.0); classification auto-detects the pooling runner from the model architecture (the old LLM(..., task="classify") arg was removed → use LLM(model)).

3. Dependencies

Include custom PyPI indexes for vLLM and FlashInfer:

# [[tool.uv.index]]
# url = "https://flashinfer.ai/whl/cu126/torch2.6"
# 
# [[tool.uv.index]]
# url = "https://wheels.vllm.ai/nightly"

Current Scripts

  1. classify-dataset.py: BERT-style text classification
    • Uses vLLM's classify task
    • Supports batch processing with configurable size
    • Automatically extracts label mappings from model config

Future Scripts

Potential additions:

  • Text generation with vLLM
  • Embedding generation using sentence transformers
  • Multi-modal inference
  • Structured output generation

Testing

Local testing requires GPU. For scripts without local GPU access:

  1. Use HF Jobs with small test datasets
  2. Verify script runs without syntax errors: python -m py_compile script.py
  3. Check dependencies resolve: uv pip compile

Performance Considerations

  • Default batch size: 10,000 for local, up to 100,000 for HF Jobs
  • L4 GPUs are cost-effective for classification
  • Monitor GPU memory usage and adjust batch sizes accordingly