v0.1.0

🚀 What is new in QuTLASS v0.1.0:

• Support for sm_100 GPUs (e.g., NVIDIA B200).
• NVFP4 Microscaling:
  • Full W4A4 quantization support.
  • Online rotations:
    • Fused transform + quantization + scale computation.
    • Rotation matrices loaded at runtime, allowing any transformation to be applied.
  • NVFP4 Matmul Kernels:
    • CUTLASS-backed NVFP4:NVFP4 with block-scale reordering.
  • Quantization:
    • Abs-Max supported.
• Multiple rotation sizes (16/32/64/128) supported for both MXFP4 and NVFP4.

v0.0.1

Core features of QuTLASS v0.0.1:

• MXFP4 microscaling support, with
• Weight and Activation quantization (W4A4)
• Online rotations: fused kernel for Hadamard tranforms, quantization, and scale computation.
  • Hadamard sizes matching the microscaling group sizes (i.e., 32 for MXFP4).
  • Compatible with any rotation matrix defined, as they are loaded in runtime.
• Multiple quantization schemes:
  • Quartet (i.e., Quest-like).
  • Abs-Max.
• Matmul kernels:
  • CUTLASS-backed MXFP4:MXFP4 kernel with block-scale reordering.
  • Prototype kernel for small batch sizes (no reordering required).
• Transformers Integration (PR #38696)