[Fix] Disable FlashInfer allreduce fusion under deterministic inference

python/sglang/srt/server_args.py

@@ -883,6 +883,12 @@ def __post_init__(self):
         # 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()
 
@@ -4020,6 +4026,12 @@ def _handle_deterministic_inference(self):
                 )
                 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(