This is meta-llama/Llama-3.1-8B-Instruct quantized with LLM Compressor in 4-bit (NVFP4), weights and activations. The calibration step used 512 samples of up to 2048 tokens, chat template applied, from HuggingFaceH4/ultrachat_200k.
The quantization has been done, tested, and evaluated by The Kaitchup. The model is compatible with vLLM. Use a Blackwell GPU to get >2x throughput.
More details in this article: NVFP4: Same Accuracy with 2.3x Higher Throughput for 4-Bit LLMs
- Developed by: The Kaitchup
- License: Apache 2.0 license
How to Support My Work
Subscribe to The Kaitchup. Or, for a one-time contribution, here is my ko-fi link: https://ko-fi.com/bnjmn_marie
This helps me a lot to continue quantizing and evaluating models for free.