Skip to content

Latest commit

 

History

History
26 lines (20 loc) · 684 Bytes

File metadata and controls

26 lines (20 loc) · 684 Bytes

Changelog

v0.0.2 [dev]

  • Changed
    • [BREAKING CHANGE] rewind validation dataloader each step
  • Added
    • Weight De-Oscillation for FP4
    • AMDFP4 (ue5m3 scales) support
  • Fixed
    • NVFP4 triton kernels without TRITON_ALLOW_NON_CONSTEXPR_GLOBALS=1

v0.0.1 [Jun 4, 2026]

Initial version of Low-Precision-Training with MXFP4/NVFP4 data types.

  • Added
    • MXFP4/NVFP4 Linear/GroupedMM Triton kernels
    • 1D/2D block quantization
    • Randomized Hadamard Transform (RHT)
    • Stochastic Rounding (SR)
    • Differential Gradient Estimation (DGE)
    • Two-Level Scaling
      • blockwise (128-dim macro-block for MXFP4)
      • tensorwise (for NVFP4)