[MoE] Disable FlashInfer fused finalize by default for numerical accuracy (#40105)

This commit is contained in:
Ziang Li
2026-09-18 00:15:57 -07:00
committed by GitHub
parent 956d414dad
commit c46bf5e990
2 changed files with 3 additions and 3 deletions
@@ -816,8 +816,8 @@ SGLang supports various environment variables that can be used to configure its
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_FLASHINFER_MOE_FUSED_FINALIZE</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use FlashInfer's fused atomic CUTLASS and CuTe DSL MoE finalize for best performance. Deterministic inference overrides this to <code>false</code>.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>true</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use FlashInfer's fused atomic CUTLASS and CuTe DSL MoE finalize. Disabled by default for better numerical accuracy. Set to <code>1</code> to opt in for more aggressive performance. Deterministic inference overrides this to <code>false</code>.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>FLASHINFER_NVFP4_4OVER6</code></td>
+1 -1
View File
@@ -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)