daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
h3-blackwell-runtime/research/nvfp4_fused_variants/README.md
Daniel Maddern ea14b98544 Preserve Vortex performance research history
2026-08-25 20:30:22 +07:00

992 B
Raw Blame History

NVFP4 Fused Variant Recovery

This capsule preserves the modulation-to-pack and SwiGLU-to-FC2-pack variants.

  • patches/0001-python-wrappers.patch: NVFP4 Python wrappers and opt-in linear methods from the current worktree.
  • patches/0002-block-call-sites.patch: current block call sites and fallback rules. The patch also includes related fused elementwise call sites and is explicitly not a kernel-only patch.
  • patches/0003-modulation-validator.patch: standalone producer validator.
  • patches/0004-swiglu-validator.patch: standalone producer validator.
  • patches/0005-elementwise-support.patch: standalone untracked Triton modulation and residual-gate support required by the block call-site patch.
  • ../shared_cuda_experimental_worktree/patches/0001-shared-cuda-experiments.patch: CUDA producer implementations and bindings; mixed with other experiments.

These artifacts preserve source only. They do not enable environment flags or change runtime/build defaults.