Unify full→SWA index translation in init_forward_metadata; drop pool caches - #27091

Merged

ch-wan merged 9 commits into

mainfrom

cheng/swa-translate-int64

Jun 3, 2026

Commits on Jun 3, 2026

fix: always use int64 for full-to-swa / hisparse index translation

ch-wan
and
claude
committed
2594187
refactor: compute full→SWA translation in init_forward_metadata; drop pool caches

ch-wan
and
claude
committed
62c6033
fix(dsv4): pass real out_cache_loc into draft-extend cuda-graph metadata

ch-wan
and
claude
committed
e1936af
fix(fa3): convert SWA page tables to int32 in flashattention backend

ch-wan
and
claude
committed
147c8a9
fix: convert SWA/hisparse page indices to int32 for int32-only attention kernels

ch-wan
and
claude
committed
629248b
fix(dsv4): compute swa_out_cache_loc in init_forward_metadata_in_graph

ch-wan
and
claude
committed
bdf57d0
fix(spec-v2): init draft-extend attn metadata after DP padding

ch-wan
and
claude
committed
3ea4b29
fix(dsv4): resolve swa_out_cache_loc via cached-or-translate helper

ch-wan
and
claude
committed
4c3c8b8
Revert "fix(spec-v2): init draft-extend attn metadata after DP padding"

ch-wan
and
claude
committed
0121ce3