21 lines
1 KiB
Markdown
21 lines
1 KiB
Markdown
# Shared CUDA Experimental Worktree Recovery
|
|
|
|
This capsule preserves the current worktree-versus-HEAD CUDA extension changes.
|
|
The patch is intentionally not described as isolated: the same translation units
|
|
contain several requested experiments and shared helpers.
|
|
|
|
Source baseline: `bee998d` (`HEAD` at recovery time).
|
|
|
|
`patches/0001-shared-cuda-experiments.patch` contains:
|
|
|
|
- Sage2 P1 fused Q/K RMSNorm, split-half RoPE, and Q INT8 preparation.
|
|
- Sage2 P2 direct V scale, E4M3 quantization, padding, and permutation.
|
|
- NVFP4 modulation-to-pack and SwiGLU-to-FC2-pack fused producers.
|
|
- Generalized physical block-scale swizzle and approximate reciprocal encoding.
|
|
- Caller-owned NVFP4 output buffers used by bounded-ring attempts.
|
|
- QKV split/layout kernels, including the vectorized 16-byte store path.
|
|
|
|
The patch also contains the shared Python wrappers required to express those
|
|
entry points. It must not be applied wholesale to production without splitting
|
|
and revalidation. Nothing in this directory is connected to runtime dispatch or
|
|
build defaults.
|