Gahow Wang a67753f516 softmax: cap block size at 512 threads ...

launch_softmax_{f32,bf16} clamped block to 1024 threads when cols was
larger. Halving the ceiling to 512 keeps two blocks per SM resident on
the large vocab kernels that dominate speculative verify workloads
without changing rows/block indexing, and never exceeds cols.

2026-07-01 14:16:32 +08:00

..

activation

gpt-oss: drop debug syncs from forward; GPU broadcast bias-add

2026-06-12 17:02:59 +08:00

attention

cuda: deterministic BF16 gemv + paged attention reductions

2026-07-01 14:16:28 +08:00

embedding

kernels/cuda: paged-attention kernel, dispatch, pinned host memory

2026-05-28 19:58:36 +08:00

gemm

cuda: deterministic BF16 gemv + paged attention reductions

2026-07-01 14:16:28 +08:00

moe

cuda: fix int32 overflow in MoE dense kernels; surface launch errors in release

2026-07-01 12:37:21 +08:00

normalization

kernels/cuda: paged-attention kernel, dispatch, pinned host memory

2026-05-28 19:58:36 +08:00

quantization

quantization: MXFP4 W4A16 expert weights (memory-optimization foundation)

2026-06-12 15:01:42 +08:00

reduce

softmax: cap block size at 512 threads

2026-07-01 14:16:32 +08:00

test

phase 0+1: project scaffold + xserv-cuda crate

2026-05-21 18:40:22 +08:00

common.cuh

cuda: fix int32 overflow in MoE dense kernels; surface launch errors in release

2026-07-01 12:37:21 +08:00