mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-17 19:40:04 +00:00
Decode tiles for split-KV hdim=64: bm0=16/1-warp and bm0=32/2-warp. Fix num_splits to use num_heads_kv (not num_heads_q) and target 4x SMs. Performance unchanged (0.056ms) because: 1. Split+combine overhead dominates for short KV (31 pages) 2. Triton 3D's single-kernel split avoids combine kernel entirely Made-with: Cursor