From e9a15b95da4463a6558760c9526893220ec6b2c8 Mon Sep 17 00:00:00 2001
From: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
Date: Sun, 10 May 2026 20:04:52 -0500
Subject: [PATCH] [Fix] Disable FlashInfer allreduce fusion under deterministic
 inference (#24629)

---
 python/sglang/srt/server_args.py | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index a2007958c..91631cc3f 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -890,6 +890,12 @@ class ServerArgs:
         # Handle memory-related, chunked prefill, and CUDA graph batch size configurations.
         self._handle_gpu_memory_settings(gpu_mem)
 
+        # enforce_disable_flashinfer_allreduce_fusion must be set before
+        # _handle_model_specific_adjustments, which auto-enables the fusion
+        # for several SM90/SM100 MoE arches.
+        if self.enable_deterministic_inference:
+            self.enforce_disable_flashinfer_allreduce_fusion = True
+
         # Apply model-specific adjustments.
         self._handle_model_specific_adjustments()
 
@@ -4078,6 +4084,12 @@ class ServerArgs:
                 )
                 self.enable_aiter_allreduce_fusion = False
 
+            if self.enable_flashinfer_allreduce_fusion:
+                logger.warning(
+                    "Disable --enable-flashinfer-allreduce-fusion because deterministic inference is enabled."
+                )
+                self.enable_flashinfer_allreduce_fusion = False
+
             # Check sampling backend
             self.sampling_backend = "pytorch"
             logger.warning(