[Perf] Use FA4 by default for MiMo on SM100 (#41886)

Co-authored-by: lianghougong <lianghouhong@xiaohongshu.com>
This commit is contained in:
leonardHONG
2026-10-01 09:05:12 +08:00
committed by GitHub
co-authored by lianghougong
parent b37e6f79fc
commit a702970ecb
2 changed files with 32 additions and 6 deletions
@@ -5,6 +5,7 @@ from typing import Any, Dict
from sglang.srt.arg_groups.model_override_base import (
_register_for,
is_attention_backend_not_set,
resolving_view,
)
from sglang.srt.runtime_context import get_platform
@@ -22,6 +23,10 @@ def _mimo_v2_overrides(server_args: Any, hf_config: Any) -> dict:
logger.info("Enable multi-layer EAGLE speculative decoding for MiMoV2 model.")
overrides["enable_multi_layer_eagle"] = True
if get_platform().is_sm100 and is_attention_backend_not_set(cfg):
overrides["attention_backend"] = "fa4"
logger.info("MiMoV2 on SM100: attention_backend=fa4.")
# On Blackwell "auto" falls through to the triton fused-MoE runner, ~12%
# slower at bs=1 decode. FP4 checkpoints use flashinfer_mxfp4 instead.
if (