SDPA decode perf improvements for qwen-3.5-35B-A3B#18759

Draft

digantdesai wants to merge 5 commits intomainfrom

digantdesai/sdpa-bench-and-perf-stats

Commits on Apr 6, 2026

[aoti-cuda] Add SDPA benchmarking script with qwen-3.5-35B-A3B shapes
digantdesai
committed

Commits on Apr 9, 2026

Print structured perf stats in Qwen3.5 MoE runner
digantdesai
committed
Add split-K decode SDPA as standalone triton_op
digantdesai
committed
Use max example seq_len when exporting Qwen3.5 MoE
digantdesai
committed
Add torch.cond split-K decode dispatch to Qwen3.5 MoE attention

digantdesai
authored and
digantdesai
committed