CoolFace
Modelpublic

ilessio-aiflowlab/Qwen2.5-Coder-14B-Instruct-NVFP4-anima

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes702downloads
Model Card

<p align="center"> <img alt="NVFP4" src="https://img.shields.io/badge/quant-NVFP4-FF3B00?style=for-the-badge&labelColor=050505"> <img alt="engine" src="https://img.shields.io/badge/engine-vLLM0.23-FF3B00?style=for-the-badge&labelColor=050505"> <img alt="hardware" src="https://img.shields.io/badge/JetsonAGXThor-sm_110a-76B900?style=for-the-badge&logo=nvidia&logoColor=white&labelColor=050505"> </p>

Qwen2.5-Coder-14B-Instruct — NVFP4

4-bit (NVFP4) quantization of [`Qwen/Qwen2.5-Coder-14B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct), produced with NVIDIA TensorRT Model Optimizer and packaged for fast inference on the NVIDIA Jetson AGX Thor (Blackwell, compute capability sm_110a).

Built with **anima-thor-ui** — RobotFlow Labs' open control plane + latest-vLLM image for the Jetson Thor.

✨ Why this build

  • —~3.5× smaller than the bf16 source — fits comfortably in the Thor's 128 GB unified memory with room for a large KV cache.
  • —Memory-bandwidth-optimal decode — NVFP4 moves ~0.55 bytes/param, roughly halving the bytes read per token versus fp8, which is what sets decode speed on bandwidth-bound edge GPUs.
  • —Drop-in OpenAI / Anthropic serving via the anima-vllm:thor-latest engine (vLLM 0.23 · PyTorch 2.11 · CUDA 13), or any vLLM build with NVFP4 (compressed-tensors) support.

🚀 Usage (vLLM)

bash
vllm serve ilessio-aiflowlab/Qwen2.5-Coder-14B-Instruct-NVFP4-anima \
  --trust-remote-code --attention-backend TRITON_ATTN \
  --gpu-memory-utilization 0.70 --kv-cache-dtype fp8 --max-model-len 32768
python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
r = client.chat.completions.create(
    model="ilessio-aiflowlab/Qwen2.5-Coder-14B-Instruct-NVFP4-anima",
    messages=[{"role": "user", "content": "Write a Python function to check if a string is a palindrome."}],
)
print(r.choices[0].message.content)

On a Jetson use --runtime nvidia (not --gpus all). The first request JIT-compiles attention (~30–60 s), then it's cached.

🔬 Quantization details

MethodNVFP4 post-training quantization (NVIDIA TensorRT Model Optimizer, NVFP4_DEFAULT_CFG)
Formatcompressed-tensors (NVFP4 weights + per-group scales; fp8 KV-cache quant)
Calibration64 diverse prompts (reasoning, code, translation, general)
Produced onNVIDIA Jetson AGX Thor · sm_110a · CUDA 13 · PyTorch 2.11 · via anima-thor-ui
Base model`Qwen/Qwen2.5-Coder-14B-Instruct`

🎯 Intended use & limitations

General-purpose inference on NVFP4-capable hardware (Blackwell and newer). Quality closely tracks the base model; as with any 4-bit PTQ, expect small deviations on the most numerically sensitive tasks. Inherits the base model's license, capabilities, biases, and intended-use terms — review the base model card. This is an independent community quantization, not affiliated with or endorsed by the base-model authors or NVIDIA.

📄 License & attribution

Released under apache-2.0 (subject to the base model's license). NVIDIA, CUDA, Jetson, and Blackwell are trademarks of NVIDIA Corporation. Quantized by [RobotFlow Labs / AIFLOW LABS](https://github.com/RobotFlow-Labs) for the ANIMA edge-AI stack using anima-thor-ui. ⭐ the repo if this is useful.