[diffusion] feat: performance_mode=speed enables torch.compile by default (#30016)

python/sglang/multimodal_gen/runtime/server_args_auto_tune.py

@@ -73,6 +73,17 @@ class ServerArgsAutoTuner:
 
         if args.performance_mode == "speed":
             logger.info("Applying performance_mode=speed")
+            if not args.enable_torch_compile and not args.is_arg_explicitly_set(
+                "enable_torch_compile"
+            ):
+                # speed means fastest: compile by default. An explicit
+                # --enable-torch-compile false still wins (e.g. models where
+                # compile measures slower, like short-step Z-Image runs).
+                args.enable_torch_compile = True
+                logger.info(
+                    "performance_mode=speed enables torch.compile "
+                    "(pass --enable-torch-compile false to opt out)"
+                )
             if args.num_gpus >= 2 and self._can_apply_fsdp_policy(
                 require_memory_headroom=False
             ):

python/sglang/multimodal_gen/test/unit/test_server_args.py

@@ -1477,6 +1477,40 @@ class TestOffloadDefaults(unittest.TestCase):
         self.assertFalse(args.text_encoder_cpu_offload)
         self.assertFalse(args.image_encoder_cpu_offload)
 
+    def test_speed_mode_enables_torch_compile_by_default(self):
+        args = self._from_dict_with_pipeline_config(
+            QwenImagePipelineConfig(),
+            kwargs={
+                "model_path": "Qwen/Qwen-Image",
+                "performance_mode": "speed",
+            },
+        )
+
+        self.assertTrue(args.enable_torch_compile)
+
+    def test_speed_mode_preserves_explicit_torch_compile_off(self):
+        args = self._from_dict_with_pipeline_config(
+            QwenImagePipelineConfig(),
+            kwargs={
+                "model_path": "Qwen/Qwen-Image",
+                "performance_mode": "speed",
+                "enable_torch_compile": False,
+            },
+        )
+
+        self.assertFalse(args.enable_torch_compile)
+
+    def test_auto_mode_leaves_torch_compile_off(self):
+        args = self._from_dict_with_pipeline_config(
+            QwenImagePipelineConfig(),
+            kwargs={
+                "model_path": "Qwen/Qwen-Image",
+                "performance_mode": "auto",
+            },
+        )
+
+        self.assertFalse(args.enable_torch_compile)
+
     def test_memory_mode_wan_uses_layerwise_offload(self):
         args = self._from_dict_with_pipeline_config(
             WanT2V480PConfig(),