debug qwen tts

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

vllm_omni/core/sched/omni_generation_scheduler.py

@@ -111,6 +111,8 @@ class OmniGenerationScheduler(VLLMScheduler):
             req_to_new_blocks[request.request_id] = new_blocks
             num_scheduled_tokens[request.request_id] = num_new_tokens
             cached_prompt_token_ids[request.request_id] = request.prompt_token_ids
+            if request.num_cached_tokens < 0:
+                request.num_cached_tokens = num_computed_tokens
             token_budget -= num_new_tokens
             scheduled_running_reqs.append(request)
             req_index += 1
@@ -177,6 +179,8 @@ class OmniGenerationScheduler(VLLMScheduler):
 
             req_to_new_blocks[request.request_id] = new_blocks
             num_scheduled_tokens[request.request_id] = num_new_tokens
+            if request.num_cached_tokens < 0:
+                request.num_cached_tokens = 0
             token_budget -= num_new_tokens
             scheduled_new_reqs.append(request)
 

vllm_omni/distributed/omni_connectors/connectors/shm_connector.py

@@ -48,7 +48,7 @@ class SharedMemoryConnector(OmniConnectorBase):
             payload = self.serialize_obj(data)
             size = len(payload)
 
-            if size > self.threshold:
+            if True:
                 # Use Shared Memory
                 lock_file = f"/dev/shm/shm_{put_key}_lockfile.lock"
                 with open(lock_file, "w") as lockf:

vllm_omni/distributed/omni_connectors/utils/serialization.py

@@ -145,7 +145,7 @@ class OmniMsgpackEncoder:
             "encoder_prompt": obj.encoder_prompt,
             "encoder_prompt_token_ids": obj.encoder_prompt_token_ids,
             "num_cached_tokens": obj.num_cached_tokens,
-            "multi_modal_placeholders": obj.multi_modal_placeholders,
+            "multi_modal_placeholders": getattr(obj, "multi_modal_placeholders", None),
             "kv_transfer_params": obj.kv_transfer_params,
         }
         # Handle dynamically added multimodal_output attribute
@@ -301,15 +301,22 @@ class OmniMsgpackDecoder:
         """Decode dict to RequestOutput.
 
         RequestOutput is not a dataclass, so msgspec.convert doesn't work.
-        We construct it manually, passing all known fields via **kwargs.
+        We construct it manually using only the known __init__ parameters to
+        avoid triggering the "Ignoring extra arguments" warning in vllm.
+        Fields that are not part of RequestOutput.__init__ (e.g.
+        multi_modal_placeholders, multimodal_output) are extracted first and
+        then restored as dynamic attributes after construction.
         """
-        # Extract multimodal_output before constructing (it's dynamically added)
+        # Extract dynamically-added / non-init fields before constructing so
+        # they are not passed as unknown **kwargs to RequestOutput.__init__.
         mm_output = obj.pop("multimodal_output", None)
+        multi_modal_placeholders = obj.pop("multi_modal_placeholders", None)
 
-        # RequestOutput.__init__ accepts **kwargs for forward compatibility
         ro = RequestOutput(**obj)
 
-        # Restore dynamically added multimodal_output attribute
+        # Restore dynamic attributes that are not part of __init__.
+        if multi_modal_placeholders is not None:
+            setattr(ro, "multi_modal_placeholders", multi_modal_placeholders)
         if mm_output is not None:
             setattr(ro, "multimodal_output", mm_output)
         return ro

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code_predictor_vllm.py

@@ -143,6 +143,7 @@ class _LocalPredictorKVCache:
                 cursor += 1
 
         max_seq_len = int(seq_lens_i32[:num_reqs].max().item())
+        max_query_len = int(query_lens_i32[:num_reqs].max().item())
         query_start_loc_gpu = qsl.to(device=self.device)
         seq_lens_gpu = seq_lens_i32.to(device=self.device)
         block_table = self._block_table[:num_reqs].contiguous()
@@ -154,6 +155,7 @@ class _LocalPredictorKVCache:
             num_tokens=num_tokens,
             query_start_loc_gpu=query_start_loc_gpu,
             query_start_loc_cpu=qsl,
+            max_query_len=max_query_len,
             seq_lens=seq_lens_gpu,
             max_seq_len=max_seq_len,
             block_tables=[block_table],