From df8137e858233685407ec266b3a6c8b030d99f30 Mon Sep 17 00:00:00 2001
From: Sy03 <1370724210@qq.com>
Date: Sat, 4 Jul 2026 16:21:02 +0800
Subject: [PATCH] Address Qwen3-TTS payload split review comments

Signed-off-by: Sy03 <1370724210@qq.com>
---
 tests/worker/test_gpu_ar_model_runner.py       | 6 ------
 vllm_omni/entrypoints/openai/serving_speech.py | 2 +-
 vllm_omni/worker/gpu_ar_model_runner.py        | 3 ---
 3 files changed, 1 insertion(+), 10 deletions(-)

diff --git a/tests/worker/test_gpu_ar_model_runner.py b/tests/worker/test_gpu_ar_model_runner.py
index d1642a248..1b108d7fe 100644
--- a/tests/worker/test_gpu_ar_model_runner.py
+++ b/tests/worker/test_gpu_ar_model_runner.py
@@ -169,7 +169,6 @@ def test_build_omni_output_uses_snapshots_and_connector_after_accumulation(monke
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=["r2"],
-        seq_len=3,
         num_scheduled_tokens_np=torch.tensor([1, 2], dtype=torch.int32).numpy(),
         query_start_loc_cpu=torch.tensor([0, 1], dtype=torch.long),
     )
@@ -215,7 +214,6 @@ def test_build_omni_output_copies_hidden_for_partial_downstream_batch(monkeypatc
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=None,
-        seq_len=6,
         num_scheduled_tokens_np=np.array([1, 2, 3], dtype=np.int32),
         query_start_loc_cpu=torch.tensor([0, 1, 3], dtype=torch.long),
     )
@@ -269,7 +267,6 @@ def test_process_additional_information_uses_snapshot_request_order(monkeypatch)
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=None,
-        seq_len=3,
         num_scheduled_tokens_np=torch.tensor([1, 2], dtype=torch.int32).numpy(),
         query_start_loc_cpu=torch.tensor([0, 1], dtype=torch.long),
     )
@@ -330,7 +327,6 @@ def test_build_omni_output_skips_hidden_when_model_opts_out(monkeypatch):
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=None,
-        seq_len=2,
         num_scheduled_tokens_np=np.array([2], dtype=np.int32),
         query_start_loc_cpu=torch.tensor([0], dtype=torch.long),
     )
@@ -377,7 +373,6 @@ def test_build_omni_output_splits_mm_by_scheduled_tokens_when_hidden_is_tail_onl
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=None,
-        seq_len=1,
         num_scheduled_tokens_np=np.array([1, 1, 1], dtype=np.int32),
         query_start_loc_cpu=torch.tensor([0, 1, 2], dtype=torch.long),
     )
@@ -535,7 +530,6 @@ def test_build_omni_output_falls_back_to_mm_cpu_without_prefix_merge(monkeypatch
         ec_connector_output=None,
         cudagraph_stats=None,
         kv_extracted_req_ids=None,
-        seq_len=2,
         num_scheduled_tokens_np=np.array([1, 1], dtype=np.int32),
         query_start_loc_cpu=torch.tensor([0, 1], dtype=torch.long),
     )
diff --git a/vllm_omni/entrypoints/openai/serving_speech.py b/vllm_omni/entrypoints/openai/serving_speech.py
index 862c57e0f..808787297 100644
--- a/vllm_omni/entrypoints/openai/serving_speech.py
+++ b/vllm_omni/entrypoints/openai/serving_speech.py
@@ -3504,7 +3504,7 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
         # keep the DELTA coercion they stream with today.
         # list() makes a copy to avoid mutating the params.
         sampling_params_list = list(self.engine_client.default_sampling_params_list)
-        async_chunk = getattr(self.engine_client.model_config, "async_chunk", True)
+        async_chunk = getattr(self.model_config, "async_chunk", True)
         qwen3_full_payload = self._tts_model_type == "qwen3_tts" and not bool(async_chunk)
         is_streaming_request = request.is_streaming() and not qwen3_full_payload
         sampling_params_list = coerce_param_message_types(sampling_params_list, is_streaming_request)
diff --git a/vllm_omni/worker/gpu_ar_model_runner.py b/vllm_omni/worker/gpu_ar_model_runner.py
index 2a49f3862..b843ed170 100644
--- a/vllm_omni/worker/gpu_ar_model_runner.py
+++ b/vllm_omni/worker/gpu_ar_model_runner.py
@@ -1545,7 +1545,6 @@ class GPUARModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin):
         ec_connector_output: Any,
         cudagraph_stats: Any,
         kv_extracted_req_ids: list[str] | None,
-        seq_len: int,
         num_scheduled_tokens_np: np.ndarray,
         query_start_loc_cpu: Any,
         postprocess_already_applied: bool = False,
@@ -1743,7 +1742,6 @@ class GPUARModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin):
             slot_mappings,  # OMNI: unpack slot_mappings for drafter
         ) = self.execute_model_state
         self.execute_model_state = None
-        seq_len = hidden_states.shape[0]
 
         # Apply structured output bitmasks if present.
         if grammar_output is not None:
@@ -1910,7 +1908,6 @@ class GPUARModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin):
                     ec_connector_output=ec_connector_output,
                     cudagraph_stats=cudagraph_stats,
                     kv_extracted_req_ids=kv_extracted_req_ids,
-                    seq_len=seq_len,
                     num_scheduled_tokens_np=num_scheduled_tokens_np,
                     query_start_loc_cpu=query_start_loc_cpu,
                     postprocess_already_applied=omni_postprocess_already_applied,