[Bugfix][StepAudio2] Fix async-chunk metadata mismatch and boundary glitches (#5917)

Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>

tests/model_executor/models/step_audio2/test_step_audio2_token2wav_async_chunk.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from types import SimpleNamespace
+from unittest.mock import MagicMock
 
 import pytest
 import torch
@@ -76,8 +77,8 @@ def test_step_audio2_token2wav_async_chunk_batch_guard():
         model._forward_async_chunk(
             input_ids=torch.tensor([1, 2, 3]),
             runtime_additional_information=[
-                {"left_context_size": 0},
-                {"left_context_size": 1},
+                {"meta": {"left_context_size": 0}},
+                {"meta": {"left_context_size": 1}},
             ],
         )
 
@@ -116,7 +117,7 @@ def test_step_audio2_token2wav_async_chunk_last_chunk_resets_state(monkeypatch):
     out = model.forward(
         input_ids=torch.tensor([10, 11, 12]),
         positions=torch.tensor([0, 1, 2]),
-        runtime_additional_information=[{"left_context_size": 1}],
+        runtime_additional_information=[{"meta": {"left_context_size": 1}}],
     )
 
     audio_list = out.multimodal_outputs["model_outputs"]
@@ -126,12 +127,24 @@ def test_step_audio2_token2wav_async_chunk_last_chunk_resets_state(monkeypatch):
     assert calls == {"setup": 1, "stream": 1, "reset": 1}
 
 
+def test_step_audio2_token2wav_routes_non_final_chunk_to_async_forward():
+    model = _build_model()
+    model._forward_async_chunk = MagicMock()
+
+    model.forward(
+        input_ids=torch.tensor([10, 11, 12]),
+        positions=torch.tensor([0, 1, 2]),
+        runtime_additional_information=[{"meta": {"left_context_size": 0}}],
+    )
+    model._forward_async_chunk.assert_called_once()
+
+
 def test_step_audio2_token2wav_async_chunk_empty_eof_returns_zero_chunk():
     model = _build_model()
 
     out = model._forward_async_chunk(
         input_ids=torch.tensor([], dtype=torch.int64),
-        runtime_additional_information=[{"left_context_size": 1}],
+        runtime_additional_information=[{"meta": {"left_context_size": 1}}],
     )
 
     audio_list = out.multimodal_outputs["model_outputs"]