[Fix] deterministic: Disable mixed chunked prefill

python/sglang/srt/arg_groups/attention_hook.py

@@ -606,6 +606,18 @@ def handle_deterministic_inference(server_args: Any):
                 enable_aiter_allreduce_fusion=False,
             )
 
+        # A mixed batch runs its decode tokens through the prefill (extend) kernels,
+        # so a decode step's result would depend on whether a prefill shares the batch.
+        if cfg.enable_mixed_chunk:
+            logger.warning(
+                "Disable --enable-mixed-chunk because deterministic inference is enabled."
+            )
+            declare_resolution(
+                server_args,
+                "_handle_deterministic_inference",
+                enable_mixed_chunk=False,
+            )
+
         # Moved to the resolution pipeline (arg_groups/overrides.py:
         # _deterministic_allreduce_fusion_disable), invoked here at its
         # legacy slot.

test/registered/unit/test_model_overrides.py

@@ -1799,6 +1799,23 @@ def test_deterministic_inference_forces_pytorch_sampling(self):
         )
         self.assertEqual(self._leaf("attention_backend"), deterministic_fills[0])
 
+    def test_deterministic_inference_disables_mixed_chunk(self):
+        """A mixed batch runs decode tokens through the extend kernels, so deterministic
+        mode must not put prefill and decode tokens in one batch."""
+        sa = self._construct(
+            "LlamaForCausalLM",
+            "llama",
+            enable_deterministic_inference=True,
+            enable_mixed_chunk=True,
+        )
+        self.assertFalse(resolution_result(sa, "enable_mixed_chunk"))
+        self.assertTrue(
+            resolution_result(
+                self._construct("LlamaForCausalLM", "llama", enable_mixed_chunk=True),
+                "enable_mixed_chunk",
+            )
+        )
+
     def test_deterministic_incompatible_backend_raises(self):
         from sglang.srt.arg_groups.overrides import (
             ResolvedView,