[CK_TILE] fix(fmha): support >2GB KV cache in batch prefill via template dispatch#6653

Merged

Jeff-Huang merged 17 commits into

users/jeff-huang/ck/batch-prefill-fix-4bg-overflow

Apr 23, 2026

Commits on Apr 22, 2026

fix(fmha): global_load_lds flat addressing for >4GB KV cache (page_size < kN0)
Jeff-Huang
committed
fix(fmha): V flat 64-bit load for >4GB KV cache (page_size < kN0)
Jeff-Huang
committed
fix(fmha): double-buffer v_physical_pages for flat load pipeline sync
Jeff-Huang
committed
feat(fmha): template dispatch for >4GB KV cache in batch prefill
Jeff-Huang
committed
refactor(fmha): unify tile_scatter_gather to two-mode design (SRD/Global load)
Jeff-Huang
committed
feat(fmha): add CDNA3+ arch guards for global_load_lds in batch prefill
Jeff-Huang
committed
fix(fmha): limit SRD num_records to page_stride after rebase (gfx950 fix)
Jeff-Huang
committed
cleanup: remove unused wave_reduce_min from utility.hpp
Jeff-Huang
committed
fix(buffer): use dependent assertion for unsupported num_dwords in async_global_load_lds_dwordxn
Jeff-Huang
committed
refactor(fmha): tighten batch prefill SRD types and document 32-bit voffset
Jeff-Huang
committed
docs(fmha): correct >2GB threshold wording across batch prefill
Jeff-Huang
committed
refactor(fmha): move use_64bit_load decision into auto-gen API dispatcher
Jeff-Huang
committed
refactor(fmha): unify kUse64BitLoad/kUseFlatLoad → kUseGlobalLoad
Jeff-Huang
committed
fix(fmha): use __builtin_assume + restore batch prefill review polish
Jeff-Huang
committed
refactor(fmha): batch prefill review polish — assert helper + setter guards
Jeff-Huang
committed

Commits on Apr 23, 2026