h3-blackwell-runtime/research/sage2_direct_vprep/README.md

16 lines
747 B
Markdown
Raw Normal View History

# Sage2 P2 Direct V Preparation Recovery
Recovered implementation sources are preserved without production integration.
- `patches/0001-vprep-validation.patch`: standalone untracked boundary-length
validator.
- `patches/0002-scheduler-decomposition.patch`: standalone untracked Sage2
decomposition harness that exercises direct V preparation.
- `../shared_cuda_experimental_worktree/patches/0001-shared-cuda-experiments.patch`:
three-stage CUDA implementation and binding. The shared patch is intentionally
mixed and must not be treated as an isolated P2 patch.
The candidate emitted Sage2's padded/permuted E4M3 V and FP32 channel scales
directly from projection-strided NHD BF16 V. It was rejected at its isolated
absolute-time gate.