From c46bf5e990bdd99e2c200214b04683022100e4df Mon Sep 17 00:00:00 2001 From: Ziang Li Date: Fri, 18 Sep 2026 00:15:57 -0700 Subject: [PATCH] [MoE] Disable FlashInfer fused finalize by default for numerical accuracy (#40105) --- docs/docs/references/environment_variables.mdx | 4 ++-- python/sglang/srt/environ.py | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/docs/docs/references/environment_variables.mdx b/docs/docs/references/environment_variables.mdx index c00648b87..84e05966c 100644 --- a/docs/docs/references/environment_variables.mdx +++ b/docs/docs/references/environment_variables.mdx @@ -816,8 +816,8 @@ SGLang supports various environment variables that can be used to configure its SGLANG_FLASHINFER_MOE_FUSED_FINALIZE - Use FlashInfer's fused atomic CUTLASS and CuTe DSL MoE finalize for best performance. Deterministic inference overrides this to false. - true + Use FlashInfer's fused atomic CUTLASS and CuTe DSL MoE finalize. Disabled by default for better numerical accuracy. Set to 1 to opt in for more aggressive performance. Deterministic inference overrides this to false. + false FLASHINFER_NVFP4_4OVER6 diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py index a84f362fa..43cfca0a4 100644 --- a/python/sglang/srt/environ.py +++ b/python/sglang/srt/environ.py @@ -1044,7 +1044,7 @@ class Envs: # token count. SGLANG_TRTLLM_MOE_PDL_MAX_TOKENS = EnvInt(8192) # Use FlashInfer's fused atomic CUTLASS/CuTe DSL MoE finalize. - SGLANG_FLASHINFER_MOE_FUSED_FINALIZE = EnvBool(True) + SGLANG_FLASHINFER_MOE_FUSED_FINALIZE = EnvBool(False) # Master switch for the experimental TRT-LLM LoRA fast path; when OFF (default) every # fine-grained opt switch reads False, keeping non-experimental paths byte-identical. SGLANG_EXPERIMENTAL_LORA_OPTI = EnvBool(False)