fix: default Inkling to fa4 on Hopper (#35292)

This commit is contained in:
Shuwen Wang
2026-10-11 20:57:42 +08:00
committed by GitHub
parent 586d98cc15
commit d26b6a5a9c
@@ -58,7 +58,7 @@ def _inkling_overrides(server_args: Any, hf_config: Any) -> dict:
# handle_mamba_radix_cache rejects extra_buffer on a model without it.
if not cfg.disable_radix_cache:
overrides["uses_mamba_radix_cache"] = True
# Inkling attention runs only on the fa4 (Blackwell) or triton backends --
# Inkling attention runs only on the fa4 (Hopper/Blackwell) or triton backends --
# models/inkling_common/attn.py asserts attention_backend in {fa4, triton}.
# The generic resolver would otherwise pick trtllm_mha (SM100) / fa3
# (Hopper), so a bare launch fails on the first attention forward. Pin a
@@ -66,7 +66,10 @@ def _inkling_overrides(server_args: Any, hf_config: Any) -> dict:
# (mirrors the MiniMax-M3 SM100 fa4-default above); an explicit
# --attention-backend / --prefill/decode-attention-backend still wins.
if is_attention_backend_not_set(cfg):
inkling_attn_backend = "fa4" if get_platform().is_sm100 else "triton"
if get_platform().is_sm90 or get_platform().is_sm100:
inkling_attn_backend = "fa4"
else:
inkling_attn_backend = "triton"
overrides["attention_backend"] = inkling_attn_backend
logger.info(
f"Use {inkling_attn_backend} as the attention backend for Inkling "