vLLM Inference Scripts

Ready-to-run UV scripts for GPU-accelerated inference using vLLM.

These scripts use UV's inline script metadata to manage dependencies. On HF Jobs they run on the vllm/vllm-openai image, which provides vLLM, PyTorch and Transformers, so the script headers leave those out. On your own Linux + CUDA machine, add vLLM explicitly: uv run --with vllm <script>.py ....

📋 Available Scripts

vlm-classify.py

Vision Language Model (VLM) image classification with structured output constraints.

Features:

• 🖼️ Process images through state-of-the-art VLMs (Qwen2-VL)
• 🎯 Structured classification using vLLM's StructuredOutputsParams
• 📐 Automatic image resizing to optimize token usage
• 💾 Memory-efficient lazy batch processing
• 🏷️ Simple CLI interface for defining classes
• 🤗 Direct integration with Hugging Face datasets

Usage:

# Basic classification
uv run --with vllm vlm-classify.py \
    username/input-dataset \
    username/output-dataset \
    --classes "document,photo,diagram,other"

# With custom prompt and image resizing
uv run --with vllm vlm-classify.py \
    username/input-dataset \
    username/output-dataset \
    --classes "index-card,manuscript,title-page,other" \
    --prompt "What type of historical document is this?" \
    --max-size 768

# Quick test with sample limit
uv run --with vllm vlm-classify.py \
    davanstrien/sloane-index-cards \
    username/test-output \
    --classes "index,content,other" \
    --max-samples 10

HF Jobs execution:

hf jobs uv run \
    --flavor l40sx1 \
    --image vllm/vllm-openai:v0.32.0 \
    -s HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/vllm/raw/main/vlm-classify.py \
    username/input-dataset \
    username/output-dataset \
    --classes "title-page,content,index,other" \
    --max-size 768

Key Parameters:

• --classes: Comma-separated list of classification categories (required)
• --prompt: Custom classification prompt (optional, auto-generated if not provided)
• --max-size: Maximum image dimension in pixels for resizing (reduces token count)
• --model: VLM model to use (default: Qwen/Qwen2-VL-7B-Instruct)
• --batch-size: Number of images to process at once (default: 8)
• --max-samples: Limit number of samples for testing

classify-dataset.py

Batch text classification using BERT-style encoder models (e.g., BERT, RoBERTa, DeBERTa, ModernBERT) with vLLM's optimized inference engine.

Note: This script is specifically for encoder-only classification models, not generative LLMs.

Features:

• 🚀 High-throughput batch processing
• 🏷️ Automatic label mapping from model config
• 📊 Confidence scores for predictions
• 🤗 Direct integration with Hugging Face Hub

Usage:

# Local execution (requires GPU)
uv run --with vllm classify-dataset.py \
    davanstrien/ModernBERT-base-is-new-arxiv-dataset \
    username/input-dataset \
    username/output-dataset \
    --inference-column text \
    --batch-size 10000

HF Jobs execution:

hf jobs uv run \
    --flavor l4x1 \
    --image vllm/vllm-openai:v0.32.0 \
    https://huggingface.co/datasets/uv-scripts/vllm/raw/main/classify-dataset.py \
    davanstrien/ModernBERT-base-is-new-arxiv-dataset \
    username/input-dataset \
    username/output-dataset \
    --inference-column text \
    --batch-size 100000

generate-responses.py

Generate responses for prompts using generative LLMs (e.g., Llama, Qwen, Mistral) with vLLM's high-performance inference engine.

Features:

• 💬 Automatic chat template application
• 📝 Support for both chat messages and plain text prompts
• 🔀 Multi-GPU tensor parallelism support
• 📏 Smart filtering for prompts exceeding context length
• 📊 Comprehensive dataset cards with generation metadata
• ⚡ HF Transfer enabled for fast model downloads
• 🎛️ Full control over sampling parameters
• 🎯 Sample limiting with --max-samples for testing

Usage:

# With chat-formatted messages (default)
uv run --with vllm generate-responses.py \
    username/input-dataset \
    username/output-dataset \
    --messages-column messages \
    --max-tokens 1024

# With plain text prompts (NEW!)
uv run --with vllm generate-responses.py \
    username/input-dataset \
    username/output-dataset \
    --prompt-column question \
    --max-tokens 1024 \
    --max-samples 100

# With custom model and parameters
uv run --with vllm generate-responses.py \
    username/input-dataset \
    username/output-dataset \
    --model-id meta-llama/Llama-3.1-8B-Instruct \
    --prompt-column text \
    --temperature 0.9 \
    --top-p 0.95 \
    --max-model-len 8192

HF Jobs execution (multi-GPU):

hf jobs uv run \
    --flavor l4x4 \
    --image vllm/vllm-openai:v0.32.0 \
    -e UV_PRERELEASE=if-necessary \
    -s HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/vllm/raw/main/generate-responses.py \
    davanstrien/cards_with_prompts \
    davanstrien/test-generated-responses \
    --model-id Qwen/Qwen3-30B-A3B-Instruct-2507 \
    --gpu-memory-utilization 0.9 \
    --max-tokens 600 \
    --max-model-len 8000

Multi-GPU Tensor Parallelism

• Auto-detects available GPUs by default
• Use --tensor-parallel-size to manually specify
• Required for models larger than single GPU memory (e.g., 30B+ models)

Handling Long Contexts

The generate-responses.py script includes smart prompt filtering:

• Default behavior: Skips prompts exceeding max_model_len
• Use --max-model-len: Limit context to reduce memory usage
• Use --no-skip-long-prompts: Fail on long prompts instead of skipping
• Skipped prompts receive empty responses and are logged

📚 About vLLM

vLLM is a high-throughput inference engine optimized for:

• Fast model serving with PagedAttention
• Efficient batch processing
• Support for various model architectures
• Seamless integration with Hugging Face models

🔧 Technical Details

UV Script Benefits

• Zero setup: Dependencies install automatically on first run
• Reproducible: Locked dependencies ensure consistent behavior
• Self-contained: Everything needed is in the script file
• Direct execution: Run from local files or URLs

Dependencies

Scripts use UV's inline metadata for dependency management, plus a [tool.hf-jobs] table that sets the image, flavor and secrets for hf jobs uv run:

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "datasets",
#     "huggingface-hub",
#     "hf-xet>= 1.1.7",
# ]
#
# [tool.hf-jobs]
# image = "vllm/vllm-openai:v0.32.0"
# python = "/usr/bin/python3"
# env = { PYTHONPATH = "/usr/local/lib/python3.12/dist-packages" }
# flavor = "a100-large"
# secrets = ["HF_TOKEN"]
# timeout = "2h"
# ///

vLLM, PyTorch, Transformers and FlashInfer come from the image: PYTHONPATH points at its packages. Do not add them to dependencies. uv would then install newer PyPI copies next to the image's, and the current flashinfer-python wheel adds a b12x module that the v0.32.0 image's vLLM imports and crashes on. For a local GPU run, use uv run --with vllm <script>.py .... hf jobs uv run reads the [tool.hf-jobs] table from hf CLI 1.32; with an older CLI, pass the header's settings as flags yourself: --image, --flavor, --timeout, --secrets HF_TOKEN, --python /usr/bin/python3 and -e PYTHONPATH=/usr/local/lib/python3.12/dist-packages.

For bleeding-edge features, use the UV_PRERELEASE=if-necessary environment variable to allow pre-release versions when needed.

Docker Image

For HF Jobs, we recommend the official vLLM Docker image: vllm/vllm-openai

This image includes:

• Pre-installed CUDA libraries
• vLLM and all dependencies
• UV package manager
• Optimized for GPU inference

Environment Variables

• HF_TOKEN: Your Hugging Face authentication token (auto-detected if logged in)
• UV_PRERELEASE=if-necessary: Allow pre-release packages when required
• HF_HUB_ENABLE_HF_TRANSFER=1: Automatically enabled for faster downloads

🔗 Resources

• vLLM Documentation
• UV Documentation
• UV Scripts Organization