QuTLASS
A CUTLASS-based quantized BLAS library for low-bit deep learning inference and training on NVIDIA Blackwell GPUs, supporting MXFP4, MXFP8 and NVFP4 microscaling formats.
QuTLASS is a low-precision kernel library built on NVIDIA CUTLASS for quantized large language model inference and training on Blackwell GPUs (sm_100 and sm_120), implementing narrow-precision microscaling routines where a scale factor covers each group of 32 elements along the inner GEMM dimension.
Version 0.2 adds a FlashInfer backend for B200 GPUs and quantization-aware training support through MXFP types, including Quartet clipping masks and prototype backward kernels for MXFP4 and MXFP8. Earlier versions added full W4A4 NVFP4 quantization with fused online rotations at multiple rotation sizes, and integrations with vLLM and Hugging Face Transformers.