minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Remove MoE prefill CUDA graph disable guard (#27841)

Browse Source
This commit is contained in:
Mohammad Miadh Angkad
2026-06-11 18:30:15 +08:00
committed by GitHub
parent 54cba63b6c
commit 6ac9f66596
1 changed files with 0 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
-16
View File
@@ -3661,22 +3661,6 @@ class ServerArgs:
self.ep_size == 1 self.ep_size == 1
), "FP8/MXFP8 Cutlass MoE is only supported with ep_size == 1" ), "FP8/MXFP8 Cutlass MoE is only supported with ep_size == 1"
# TODO(yuwei): Fix piecewise cuda graph support for bypassed topk MoE backends.
# Exception: GptOssForCausalLM wraps the entire MoE block in its own
# custom op (moe_impl), so bypassed topk is handled inside the op body.
if (
(Phase.PREFILL, "backend") not in self._cuda_graph_config_locked
and self.moe_runner_backend in ("flashinfer_trtllm", "flashinfer_mxfp4")
and self.get_model_config().hf_config.architectures[0]
!= "GptOssForCausalLM"
):
self.cuda_graph_config.prefill.backend = Backend.DISABLED
logger.info(
f"Piecewise cuda graph is disabled for MoE runner backend "
f"'{self.moe_runner_backend}' (bypassed topk is incompatible "
f"with torch.compile)."
)
def cutedsl_moe_max_num_tokens(self) -> int: def cutedsl_moe_max_num_tokens(self) -> int:
"""Largest number of tokens a single forward routes through a CuteDSL """Largest number of tokens a single forward routes through a CuteDSL
MoE layer on one (DP) rank. Single source of truth for both the MoE layer on one (DP) rank. Single source of truth for both the