[diffusion][MOVA] fix: fix task type in MOVA pipeline and shared model placement (#19489)

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

@@ -27,7 +27,7 @@ logger = init_logger(__name__)
 class MOVAPipelineConfig(PipelineConfig):
     """Configuration for MOVA (text+image -> video+audio) pipelines."""
 
-    task_type: ModelTaskType = ModelTaskType.T2V
+    task_type: ModelTaskType = ModelTaskType.I2V
 
     # Model configs
     dit_config: MOVAVideoConfig = field(default_factory=MOVAVideoConfig)

python/sglang/multimodal_gen/configs/sample/mova.py

@@ -19,6 +19,8 @@ class MOVASamplingParams(SamplingParams):
     visual_shift: float = 5.0
     audio_shift: float = 5.0
 
+    adjust_frames: bool = False
+
     negative_prompt: str = (
         "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，"
         "整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，"

python/sglang/multimodal_gen/runtime/models/bridges/mova_dual_tower.py

@@ -220,7 +220,7 @@ class ConditionalCrossAttention(nn.Module):
             head_size=self.head_dim,
             causal=False,
             softmax_scale=None,
-            is_cross_attention=True,
+            # is_cross_attention=True,
         )
 
     def forward(

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/mova.py

@@ -349,6 +349,11 @@ class MOVADenoisingStage(PipelineStage):
         self._manage_device_placement(current_model, model_to_offload, server_args)
         return current_model
 
+    def _ensure_shared_models_on_device(self, server_args: ServerArgs):
+        """Ensure shared denoising modules are on the active device when cpu offload is enabled."""
+        self._manage_device_placement(self.audio_dit, None, server_args)
+        self._manage_device_placement(self.dual_tower_bridge, None, server_args)
+
     def _apply_guidance_rescale(
         self,
         noise_pred,
@@ -378,7 +383,7 @@ class MOVADenoisingStage(PipelineStage):
     @torch.no_grad()
     def forward(self, batch: Req, server_args: ServerArgs) -> Req:
         self._maybe_compile_dits(server_args)
-        self._manage_device_placement(self.audio_dit, None, server_args)
+        self._ensure_shared_models_on_device(server_args)
 
         paired_timesteps = batch.paired_timesteps
         if paired_timesteps is None: