improve gdn mtp bf16 state perf for BS<=8 with LDG.128#3143

Draft

ameynaik-hub wants to merge 5 commits intoflashinfer-ai:mainfrom

ameynaik-hub:ameyn/wide_vec_tile_v

Commits on Apr 19, 2026

perf(gdn): fix bf16_state T=1 per-call overhead and add pool+padding to ILP kernel

ameynaik-hub
and
claude
committed

Commits on Apr 20, 2026

Merge branch 'main' into ameyn/bf16_baseline_fix
ameynaik-hub
authored
perf(gdn): add wide_vec BF16 MTP kernel, auto-dispatch, and T=2 heuristic fix

ameynaik-hub
and
claude
committed
perf(gdn): restore ILP=4 MTP launcher for small-batch BF16 MTP decode

ameynaik-hub
and
claude
committed

Commits on Apr 21, 2026

perf(gdn): extend wide_vec fast path to small batches via tile_v

ameynaik-hub
and
claude
committed