README.md

---
license: mit
library_name: transformers
pipeline_tag: image-text-to-text
base_model:
- numind/NuExtract-2.0-8B
---

# NuExtract-2.0-8B-FP8-Dynamic

## Quantization

Quantized with [llm-compressor](https://github.com/vllm-project/llm-compressor) v0.9.0.1.

We used the original [qwen2.5-vl example compression script](https://github.com/vllm-project/llm-compressor/blob/main/examples/multimodal_vision/qwen_2_5_vl_example.py) and adapted it to a [FP8-Dynamic compression recipe](https://github.com/vllm-project/llm-compressor/tree/main/examples/quantization_w8a8_fp8).

## vLLM inference

```bash
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](https://github.com/vllm-project/vllm/issues/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.