13 lines
653 B
Markdown
13 lines
653 B
Markdown
# QKV Layout, Swizzle, and Wide-Store Recovery
|
|
|
|
The recoverable source is split between the shared CUDA patch and a focused
|
|
Python dispatch patch:
|
|
|
|
- `patches/0001-strided-nhd-dispatch.patch`: current `attention.py` changes for
|
|
projection-strided NHD Sage2 and contiguous token-major output.
|
|
- `../shared_cuda_experimental_worktree/patches/0001-shared-cuda-experiments.patch`:
|
|
QKV split/layout kernels, vectorized 16-byte stores, generalized NVFP4 scale
|
|
swizzle, and associated extension binding. This patch also contains other
|
|
requested CUDA experiments and is not isolated.
|
|
|
|
No experimental dispatch or build default is added by this capsule.
|