mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 19:28:33 +00:00
Checkpoint before merging upstream CK. Adds the kIsPaged=false kernel instances (d64/d128 bf16/fp8, mask/nmask), folds kv_start into base pointers, and fixes the non-causal KV-range envelope (scan full seq_len when !FmhaMask::IsMasking instead of the causal horizon). Co-authored-by: Cursor <cursoragent@cursor.com>