4 Bit Quantization With Nf4 Data Type For Llm Weight Compression

via “quantized inference with 8-bit and mxfp4 precision”

text-generation model by undefined. 65,88,909 downloads.

Unique: Native support for mxfp4 quantization format (mixed-precision floating-point) alongside standard 8-bit integer quantization, providing fine-grained control over precision-performance tradeoffs. Integrated with vLLM's optimized CUDA kernels for quantized inference, achieving 2-3x speedup compared to naive quantization implementations.

vs others: Offers mxfp4 as middle ground between 8-bit (faster but lower quality) and full precision, whereas most open-source models only support 8-bit or require external quantization tools like GPTQ or AWQ