[Perf] gdn: Select FlashInfer prefill for page-major pools (#43171)

Co-authored-by: Caihua Li <caihua.li@bytedance.com>
This commit is contained in:
caihuali95
2026-10-18 05:37:24 -07:00
committed by GitHub
co-authored by Caihua Li
parent ba50ea3d77
commit 1d17066bad
4 changed files with 28 additions and 5 deletions
@@ -832,14 +832,15 @@ def handle_page_major_kv_layout(server_args: Any):
# stride as a free int64); helion (specializes KDA state strides 0-3
# and rejects a non-unit innermost stride); cutedsl (KDA fused sigmoid-
# gating update is stride-safe) on KDA-hybrid models only.
# - prefill: triton; flashkda (the wrapper gathers/scatters a contiguous
# per-slot copy); helion; cutedsl (kernel_h compiles h0/ht with dynamic
# - prefill: triton; flashinfer (GDN gathers/scatters contiguous state);
# flashkda (the wrapper gathers/scatters a contiguous per-slot copy);
# helion; cutedsl (kernel_h compiles h0/ht with dynamic
# int64 strides), with the same KDA-only caveat.
# - mamba (mamba2/short-conv state): triton only.
# use_mla_backend() distinguishes the KDA-hybrid family (K3/KimiLinear
# are MLA-hybrid) from GDN models (GQA-hybrid) for the KDA-only caveat.
decode_allowed = {"triton", "flashinfer"}
prefill_allowed = {"triton", "flashkda"}
prefill_allowed = {"triton", "flashkda", "flashinfer"}
if use_mla_backend(server_args):
decode_allowed.update({"cutedsl", "helion"})
prefill_allowed.update({"cutedsl", "helion"})
@@ -195,7 +195,6 @@ def flashinfer_gdn_prefill_default(model_runner: ModelRunner) -> Optional[str]:
get_exec().mamba.linear_attn_prefill_backend is not None
or get_exec().mamba.linear_attn_backend != "triton"
or get_exec().deterministic.enable_deterministic_inference
or get_memory().enable_page_major_kv_layout
or sm_major not in (9, 10)
):
return None