Pascal ebb8a1706f cuda: use the vector lightning indexer kernel on MUSA (llama/29990) ...

* cuda: stage the lightning indexer queries in head passes for MUSA

MUSA archs 21 and 22 cap static shared memory at 28 KB, and the tile
kernel staged the queries of all four heads next to the key tile for
33 KB. The queries are now staged in passes of
LIGHTNING_INDEXER_TILE_HEADS_PER_PASS heads: two on MUSA for 25 KB,
four elsewhere where the single pass folds to the previous kernel.

* cuda: use the vector lightning indexer kernel on MUSA

Address review from am17an: the tile kernel stays off MUSA, whose archs
21 and 22 cap static shared memory at 28 KB, below the 33 KB the tile
needs, so MUSA keeps the vector kernel it ran before. This replaces the
head passes, CUDA and ROCm run the merged kernel unchanged.

2026-10-06 10:38:22 +03:00

..

cmake

Enables Windows ARM64 build with MSVC cl.exe (llama/28362)

2026-10-06 10:38:06 +03:00

include

ggml : add alloc_buffer_n to buffer type interface (llama/23671)

2026-10-06 10:38:16 +03:00

src

cuda: use the vector lightning indexer kernel on MUSA (llama/29990)

2026-10-06 10:38:22 +03:00

.gitignore

…

CMakeLists.txt

CUDA: Make CCCL configurable + pin it to 3.4.3 for CI jobs (llama/29792)

2026-10-06 10:38:14 +03:00