[Bug-fix] Fix Bugs in Qwen3/Qwen2.5 Omni Rebased Support

vllm_omni/entrypoints/omni_stage.py

@@ -452,9 +452,9 @@ def filter(self, record: _logging.LogRecord) -> bool:
         max_batch_size = int(runtime_cfg.get("max_batch_size", 1) or 1)
         print(f"[Stage-{stage_id}] Max batch size: {max_batch_size}")
         batch_tasks: list[dict[str, Any]] = [task]
+        start_time = _time.time()
         if max_batch_size > 1:
             while len(batch_tasks) < max_batch_size:
-                start_time = _time.time()
                 if not in_q.empty():
                     extra = in_q.get_nowait()
                     if extra is None:

vllm_omni/worker/gpu_model_runner.py

@@ -6,6 +6,8 @@
 from vllm.distributed.parallel_state import get_pp_group
 from vllm.forward_context import BatchDescriptor, set_forward_context
 from vllm.logger import init_logger
+from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
+from vllm.model_executor.models.interfaces import supports_mrope
 from vllm.model_executor.models.interfaces_base import VllmModelForPooling
 from vllm.sampling_params import SamplingType
 from vllm.utils import LazyLoader, cdiv
@@ -33,6 +35,63 @@
 
 
 class OmniGPUModelRunner(GPUModelRunner):
+    def _init_mrope_positions(self, req_state: CachedRequestState):
+        """Initialize M-RoPE positions for multimodal inputs.
+
+        Extracts multimodal feature metadata (image grids, video grids,
+        audio features) and computes M-RoPE positions for proper positional
+        encoding of multimodal tokens.
+
+        Args:
+            req_state: Cached request state containing multimodal features
+
+        Raises:
+            AssertionError: If the model does not support M-RoPE
+        """
+        image_grid_thw = []
+        video_grid_thw = []
+        second_per_grid_ts = []
+        audio_feature_lengths = []
+        use_audio_in_video = False
+        for mm_feature in req_state.mm_features:
+            mm_item = mm_feature.data
+            if mm_item is None:
+                continue
+            mm_input = mm_item.get_data()
+            if (t := mm_input.get("image_grid_thw")) is not None:
+                image_grid_thw.append(t.tolist())
+            if (t := mm_input.get("video_grid_thw")) is not None:
+                video_grid_thw.append(t.tolist())
+            if (t := mm_input.get("second_per_grid_ts")) is not None:
+                second_per_grid_ts.append(t)
+            if (t := mm_input.get("audio_feature_lengths")) is not None:
+                audio_feature_lengths.append(t)
+            # Check for use_audio_in_video
+            use_audio_in_video_value = mm_input.get("use_audio_in_video")
+            if use_audio_in_video_value is not None:
+                use_audio_in_video = bool(use_audio_in_video_value.item())
+
+        if supports_mrope(self.model):
+            req_state.mrope_positions, req_state.mrope_position_delta = self.model.get_mrope_input_positions(
+                req_state.prompt_token_ids,
+                hf_config=self.model_config.hf_config,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                second_per_grid_ts=second_per_grid_ts,
+                audio_feature_lengths=audio_feature_lengths,
+                use_audio_in_video=use_audio_in_video,
+            )
+        else:
+            req_state.mrope_positions, req_state.mrope_position_delta = MRotaryEmbedding.get_input_positions_tensor(
+                req_state.prompt_token_ids,
+                hf_config=self.model_config.hf_config,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                second_per_grid_ts=second_per_grid_ts,
+                audio_feature_lengths=audio_feature_lengths,
+                use_audio_in_video=use_audio_in_video,
+            )
+
     def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
         """Update the cached states and the persistent batch with the scheduler
         output.