Update vLLM precompiled wheel version and refactor various components for improved functionality and error handling. Changes include modifying the get_name method in DiffusionQuantizationConfig to use class-level attributes, enhancing metadata handling in StageEngineCoreClient, and adjusting the AsyncOmni class constructor for better parameter management. Additionally, improved error handling in omni_snapshot_download and optimized initial chunk size computation in talker2code2wav_async_chunk for dynamic load adaptation.

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

vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_moe_thinker.py

@@ -155,8 +155,7 @@ class Qwen3Omni_VisionTransformer(_Qwen3Omni_VisionTransformer):
             cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
         except RuntimeError:
             logger.warning(
-                "torch.repeat_interleave not executable, "
-                "switching to vectorized searchsorted implementation."
+                "torch.repeat_interleave not executable, switching to vectorized searchsorted implementation."
             )
             repeat_counts = grid_thw_tensor[:, 0]
             values = grid_thw_tensor[:, 1] * grid_thw_tensor[:, 2]
@@ -179,9 +178,9 @@ class Qwen3Omni_VisionTransformer(_Qwen3Omni_VisionTransformer):
         max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
 
         grid_thw_np = grid_thw_tensor.cpu().numpy().astype(np.int32)
-        cu_seqlens_np = np.repeat(
-            grid_thw_np[:, 1] * grid_thw_np[:, 2], grid_thw_np[:, 0]
-        ).cumsum(axis=0, dtype=np.int32)
+        cu_seqlens_np = np.repeat(grid_thw_np[:, 1] * grid_thw_np[:, 2], grid_thw_np[:, 0]).cumsum(
+            axis=0, dtype=np.int32
+        )
         cu_seqlens_np = np.concatenate([np.zeros(1, dtype=np.int32), cu_seqlens_np])
         sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens(
             self.attn_backend,

vllm_omni/model_executor/stage_input_processors/qwen3_tts.py

@@ -116,18 +116,12 @@ def talker2code2wav_async_chunk(
             initial_chunk_size = int(entry.list_data[0])
             per_request_override = True
 
-    # Dynamic IC: cache per request so boundaries stay stable for its lifetime.
+    # Dynamic IC: recalculate each call so IC adapts to current load.
     if not per_request_override:
-        _ic_cache = getattr(transfer_manager, "_cached_ic", None)
-        if _ic_cache is None:
-            _ic_cache = {}
-            transfer_manager._cached_ic = _ic_cache
-        if request_id not in _ic_cache:
-            max_ic = max_ic_for_chunk_size(chunk_size)
-            active = sum(1 for v in transfer_manager.code_prompt_token_ids.values() if len(v) > 0)
-            capacity = getattr(transfer_manager, "scheduler_max_num_seqs", 1)
-            _ic_cache[request_id] = compute_dynamic_initial_chunk_size(active, capacity, max_ic)
-        initial_chunk_size = _ic_cache[request_id]
+        max_ic = max_ic_for_chunk_size(chunk_size)
+        active = sum(1 for v in transfer_manager.code_prompt_token_ids.values() if len(v) > 0)
+        capacity = getattr(transfer_manager, "scheduler_max_num_seqs", 1)
+        initial_chunk_size = compute_dynamic_initial_chunk_size(active, capacity, max_ic)
 
     if chunk_size <= 0 or left_context_size_config < 0 or initial_chunk_size < 0:
         raise ValueError(
@@ -145,11 +139,13 @@ def talker2code2wav_async_chunk(
         initial_chunk_size = chunk_size
     length = len(transfer_manager.code_prompt_token_ids[request_id])
 
+    finished_tensor = torch.tensor(finished, dtype=torch.bool)
+
     if length <= 0:
         if finished:
             return {
                 "code_predictor_codes": [],
-                "finished": True,
+                "finished": finished_tensor,
             }
         return None
 
@@ -197,5 +193,5 @@ def talker2code2wav_async_chunk(
     return {
         "code_predictor_codes": code_predictor_codes,
         "left_context_size": left_context_size,
-        "finished": finished,
+        "finished": finished_tensor,
     }