fix: LFM2.5 paged turbo4 on gfx803 (warp-width, sub-128 head_dim, parallel partials) - #13

Merged

kmbandy merged 3 commits into

fix/lfm25-paged-turbo4-gfx803

Jun 25, 2026

Commits on Jun 15, 2026

fix(cuda): pin turbo4 cooperative warp-collectives to WARP_SIZE (gfx803 corruption)
kmbandy
committed
62f4f99
feat(paged): support sub-128 head_dim for paged turbo cache (LFM2.5 head_dim 64)

kmbandy
and
claude
committed
fb556b1
fix(cuda): size paged-decode partials by total_q_tokens, not avg_q_len

kmbandy
and
claude
committed
93a1e6a