[diffusion] chore: make torch.compile opt-in for speed mode (#34173)

python/sglang/multimodal_gen/configs/pipeline_configs/model_deployment_config.py

@@ -25,9 +25,9 @@ class ModelDeploymentConfig:
     auto_enable_cfg_parallel: bool = True
     # degree 1 keeps CFG parallel disabled and leaves GPUs available for SP
     auto_cfg_parallel_degree_by_num_gpus: tuple[tuple[int, int], ...] = ()
-    # Let performance_mode=speed opt into torch.compile unless the model has
-    # established that the compiled path changes its numerical contract.
-    speed_mode_enable_torch_compile_by_default: bool = True
+    # torch.compile is model opt-in because it can be slower than eager for
+    # diffusion workloads dominated by already-optimized kernels
+    speed_mode_enable_torch_compile_by_default: bool = False
     supports_cfg_parallel: bool = True
 
     def get_auto_cfg_parallel_degree(self, num_gpus: int) -> int:

python/sglang/multimodal_gen/runtime/server_args/auto_tune.py

@@ -78,9 +78,7 @@ class ServerArgsAutoTuner:
                 and 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 is slower or changes the numerical contract).
+                # only models with a validated compile win opt in by default
                 args.enable_torch_compile = True
                 logger.info(
                     "performance_mode=speed enables torch.compile "

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

@@ -24,6 +24,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.ltx_2 import (
 from sglang.multimodal_gen.configs.pipeline_configs.minimax_h3 import (
     MiniMaxH3PipelineConfig,
 )
+from sglang.multimodal_gen.configs.pipeline_configs.model_deployment_config import (
+    ModelDeploymentConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.mova import MOVAPipelineConfig
 from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import (
     QwenImagePipelineConfig,
@@ -1590,7 +1593,7 @@ 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):
+    def test_speed_mode_keeps_torch_compile_off_by_default(self):
         args = self._from_dict_with_pipeline_config(
             QwenImagePipelineConfig(),
             kwargs={
@@ -1599,20 +1602,40 @@ class TestOffloadDefaults(unittest.TestCase):
             },
         )
 
-        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_speed_mode_preserves_explicit_torch_compile_setting(self):
+        for enabled in (False, True):
+            with self.subTest(enabled=enabled):
+                args = self._from_dict_with_pipeline_config(
+                    QwenImagePipelineConfig(),
+                    kwargs={
+                        "model_path": "Qwen/Qwen-Image",
+                        "performance_mode": "speed",
+                        "enable_torch_compile": enabled,
+                    },
+                )
+
+                self.assertEqual(args.enable_torch_compile, enabled)
+
+    def test_speed_mode_honors_model_torch_compile_opt_in(self):
+        with patch.object(
+            QwenImagePipelineConfig,
+            "get_model_deployment_config",
+            return_value=ModelDeploymentConfig(
+                speed_mode_enable_torch_compile_by_default=True
+            ),
+        ):
+            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_uses_minimax_h3_compile_policy(self):
         for explicit, expected in ((None, False), (True, True)):
             kwargs = {