[codex] Optimize DMD Wan auto residency on high-memory GPUs (#28780)

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

@@ -108,6 +108,22 @@ class TurboWanT2V480PConfig(WanT2V480PConfig):
     )
 
 
+@dataclass
+class TurboWanT2V1_3B480PConfig(TurboWanT2V480PConfig):
+    """Configuration for TurboWan T2V 1.3B DMD pipeline."""
+
+    def get_model_deployment_config(self) -> ModelDeploymentConfig:
+        return ModelDeploymentConfig(
+            auto_dit_layerwise_offload=True,
+            auto_disable_component_offload_min_available_memory_gb=60,
+            auto_disable_component_offload_components=(
+                "text_encoder",
+                "image_encoder",
+                "vae",
+            ),
+        )
+
+
 @dataclass
 class WanT2V720PConfig(WanT2V480PConfig):
     """Base configuration for Wan T2V 14B 720P pipeline architecture."""
@@ -183,6 +199,17 @@ class FastWan2_1_T2V_480P_Config(WanT2V480PConfig):
         default_factory=lambda: [1000, 757, 522]
     )
 
+    def get_model_deployment_config(self) -> ModelDeploymentConfig:
+        return ModelDeploymentConfig(
+            auto_dit_layerwise_offload=True,
+            auto_disable_component_offload_min_available_memory_gb=60,
+            auto_disable_component_offload_components=(
+                "text_encoder",
+                "image_encoder",
+                "vae",
+            ),
+        )
+
 
 @dataclass
 class Wan2_2_TI2V_5B_Config(WanT2V480PConfig, WanI2VCommonConfig):

python/sglang/multimodal_gen/registry.py

@@ -88,6 +88,7 @@ from sglang.multimodal_gen.configs.pipeline_configs.wan import (
     FastWan2_1_T2V_480P_Config,
     FastWan2_2_TI2V_5B_Config,
     TurboWanI2V720Config,
+    TurboWanT2V1_3B480PConfig,
     TurboWanT2V480PConfig,
     Wan2_2_I2V_A14B_Config,
     Wan2_2_T2V_A14B_Config,
@@ -674,7 +675,7 @@ def _register_configs():
     )
     register_configs(
         sampling_param_cls=WanT2V_1_3B_SamplingParams,
-        pipeline_config_cls=TurboWanT2V480PConfig,
+        pipeline_config_cls=TurboWanT2V1_3B480PConfig,
         hf_model_paths=[
             "IPostYellow/TurboWan2.1-T2V-1.3B-Diffusers",
         ],

python/sglang/multimodal_gen/runtime/server_args_auto_tune.py

@@ -95,8 +95,9 @@ class ServerArgsAutoTuner:
             return
 
         min_available_gb = self._get_min_available_device_memory_gb()
+        deployment_config = self._deployment_config()
         disable_threshold_gb = (
-            self._deployment_config().auto_disable_component_offload_min_available_memory_gb
+            deployment_config.auto_disable_component_offload_min_available_memory_gb
         )
         if (
             min_available_gb is not None
@@ -104,9 +105,7 @@ class ServerArgsAutoTuner:
             and min_available_gb >= disable_threshold_gb
         ):
             changed = []
-            components = (
-                self._deployment_config().auto_disable_component_offload_components
-            )
+            components = deployment_config.auto_disable_component_offload_components
             if (
                 args.layerwise_offload_components is not None
                 and not args.is_arg_explicitly_set("layerwise_offload_components")
@@ -387,11 +386,50 @@ class ServerArgsAutoTuner:
             for component_name, arg_name in DEFAULT_LAYERWISE_COMPONENT_ARG_NAMES
             if not args.is_arg_explicitly_set(arg_name)
         ]
+        components = self._filter_high_memory_resident_components(components)
         if self._should_auto_enable_dit_layerwise_offload():
             components.insert(0, LAYERWISE_OFFLOAD_DIT_GROUP)
             self._set_default_wan_dit_offload_prefetch_size()
         return components
 
+    def _filter_high_memory_resident_components(
+        self, components: list[str]
+    ) -> list[str]:
+        args = self.server_args
+        if args.performance_mode != "auto" or current_platform.is_cpu():
+            return components
+
+        deployment_config = self._deployment_config()
+        threshold_gb = (
+            deployment_config.auto_disable_component_offload_min_available_memory_gb
+        )
+        if threshold_gb is None:
+            return components
+
+        min_available_gb = self._get_min_available_device_memory_gb()
+        if min_available_gb is None or min_available_gb < threshold_gb:
+            return components
+
+        resident_components = set(
+            deployment_config.auto_disable_component_offload_components
+        )
+        filtered_components = [
+            component
+            for component in components
+            if component not in resident_components
+        ]
+        skipped_components = [
+            component for component in components if component in resident_components
+        ]
+        if skipped_components:
+            logger.info(
+                "Keeping default layerwise components resident for %s because minimum available memory on selected GPUs is %.2f GiB: %s",
+                args.pipeline_config.__class__.__name__,
+                min_available_gb,
+                ", ".join(skipped_components),
+            )
+        return filtered_components
+
     def _should_auto_enable_dit_layerwise_offload(self) -> bool:
         args = self.server_args