h3-blackwell-runtime/research/cute_nvfp4_ring
2026-08-25 20:30:22 +07:00
..
patches Preserve Vortex performance research history 2026-08-25 20:30:22 +07:00
README.md Preserve Vortex performance research history 2026-08-25 20:30:22 +07:00
UNRECOVERABLE.md Preserve Vortex performance research history 2026-08-25 20:30:22 +07:00

CuTe NVFP4 and Ring Attempt Recovery

The bounded-ring runtime module and validator-only CuTe experiments were already isolated as untracked files, so they are preserved as add-file patches:

  • patches/0001-cute-qkv-ring-module.patch
  • patches/0002-conversion-contract.patch
  • patches/0003-tile-producer.patch
  • patches/0004-ring-validator.patch
  • patches/0005-real-tile-validator.patch
  • patches/0006-h3-gemm-validator.patch
  • patches/0007-full-ring-validator.patch
  • patches/0008-qkv-block-gate.patch
  • patches/0009-qkv-runtime-gate.patch

The caller-owned _into CUDA producer needed by the ring is in the explicitly mixed shared CUDA patch under ../shared_cuda_experimental_worktree/patches/. The CuTe kernel itself was referenced through an external path and is not in the worktree or reachable Git objects; see UNRECOVERABLE.md.

Nothing here is registered with production dispatch or build defaults.