[NVIDIA] Integrate FlashInfer decode kernel (Blackwell) for Qwen3.5#19150

Merged

ispobock merged 7 commits intosgl-project:mainfrom

kaixih:integrate_flashinfer_gdn

Mar 18, 2026

Commits on Mar 16, 2026

feat(gdn): integrate FlashInfer decode into linear attention backend

kaixih
and
claude
committed

Commits on Mar 17, 2026

fix(gdn): remove prefill gather+transpose for FlashInfer pool path

kaixih
and
claude
committed
debug(gdn): replace FlashInfer SM100+ decode with Triton at callsite

kaixih
and
claude
committed
feat(gdn): restore FlashInfer SM100+ pool decode with dt_bias fp32 fix and add flashinfer gsm8k test

kaixih
and
claude
committed
refactor(gdn): merge FP4 Triton/FlashInfer tests and add top_k to AccuracyTestParams

kaixih
and
claude
committed
fix(gdn): raise error when flashinfer decode used with no_buffer scheduling

kaixih
and
claude
committed
fix(gdn): move no_buffer+flashinfer guard before extra_buffer validation

kaixih
and
claude
committed