NuExtract-2.0-8B-FP8-Dynamic

Quantization

Quantized with llm-compressor v0.9.0.1.

We used the original qwen2.5-vl example compression script and adapted it to a FP8-Dynamic compression recipe.

vLLM inference

docker run --rm --name 'NuExtract-2.0-8B' -e HF_TOKEN -v '/srv/cache:/root/.cache' -p 8000:8000 -e LD_LIBRARY_PATH='/lib/x86_64-linux-gnu:/usr/local/cuda/lib64' 'vllm/vllm-openai:v0.15.1-cu130' 'ig1/NuExtract-2.0-8B-FP8-Dynamic' --served-model-name 'NuExtract-2.0-8B' --trust-remote-code --limit-mm-per-prompt '{"image": 6, "video": 0}' --chat-template-content-format 'openai' --max-model-len 'auto' --kv-cache-memory-bytes '7G'

• -e LD_LIBRARY_PATH=/lib64:/usr/local/cuda/lib64 is only needed if your host have a recent driver version (with native CUDA 13.0 or 13.1). See #32373 for more info.
• Adapt /srv/cache to your liking, this will contains all cache data you want to keep for faster startup:
  • dirs like huggingface, torch, vllm, flashinfer, etc...
• To avoid eating up all the GPU VRAM, the --kv-cache-memory-bytes '7G' is set (it allows max model len by 1.02x)
  • Feel free to adjust (or remove the flag and switch back to --gpu-memory-utilization 0.9) to increase or decrease KV cache to your liking

Check original project readme for openai like chat template use in requests.