[Rebase] Rebase to vllm 0.26.0 (#5443)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Signed-off-by: tzhouam <tzhouam@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: tzhouam <tzhouam@gmail.com>

tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py

@@ -201,6 +201,7 @@ def _assert_realtime_smoke(result: dict) -> None:
 def _assert_realtime_accuracy(
     result: dict,
     whisper_model_size: str = "large-v3",
+    threshold: float = 0.8,
 ) -> None:
     """Assert that whisper transcription of audio output matches model text.
 
@@ -214,6 +215,12 @@ def _assert_realtime_accuracy(
                    variability even though audio generation was correct. large-v3
                    transcribes these clips reliably, so a failure here now points
                    at the model, not the ASR grader.
+        threshold: Minimum cosine similarity (with length penalty) required to
+                   pass. Default 0.8. Do not lower per-callsite without data:
+                   at 0.35 the assertion no longer detects real audio
+                   regressions. If a variant genuinely needs a different gate
+                   (e.g. whisper partial transcripts under async_chunk), propose
+                   it in its own PR with measurements.
     """
     final_text = (result["transcription_text"] or "").strip()
     assert final_text, "Expected non-empty transcription (model text stream)"
@@ -223,8 +230,9 @@ def _assert_realtime_accuracy(
     assert whisper_text, "Whisper returned empty string for synthesized output audio"
 
     sim = cosine_similarity_text(whisper_text.lower(), final_text.lower())
-    assert sim > 0.8, (
-        f"Output audio transcript should match model text (sim={sim:.3f}): "
+    assert sim > threshold, (
+        f"Output audio transcript should match model text (sim={sim:.3f}, "
+        f"threshold={threshold}): "
         f"whisper={whisper_text!r}, model_text={final_text!r}"
     )
 

tests/entrypoints/test_pd_disaggregation.py

@@ -305,21 +305,11 @@ def mock_get_config(monkeypatch):
     """Auto-mock get_config and related model loading functions."""
     import sys
 
-    fake_tokenizer = _ns()
-    fake_tokenizer.encode = lambda *args, **kwargs: [1, 2, 3]
-    fake_tokenizer.decode = lambda *args, **kwargs: "test"
-
-    def _mock_init_tokenizer_from_configs(model_config=None, **kwargs):
-        return fake_tokenizer
-
-    monkeypatch.setattr(
-        "vllm.transformers_utils.tokenizer.init_tokenizer_from_configs",
-        _mock_init_tokenizer_from_configs,
-        raising=False,
-    )
-    tokenizer_module_path = "vllm.transformers_utils.tokenizer"
-    if tokenizer_module_path in sys.modules:
-        setattr(sys.modules[tokenizer_module_path], "init_tokenizer_from_configs", _mock_init_tokenizer_from_configs)
+    # TODO(pd-unskip): the tokenizer mock that lived here targeted
+    # vllm.transformers_utils.tokenizer.init_tokenizer_from_configs, which was
+    # deleted in vLLM 0.26 (the raising=False setattr made it a silent no-op).
+    # When the module-level skip is lifted, mock whatever tokenizer seam the
+    # revived PD entrypoint code actually uses.
 
     def _mock_length_from_prompt_token_ids_or_embeds(prompt_token_ids=None, prompt_embeds=None):
         if prompt_token_ids is not None:
@@ -344,13 +334,6 @@ def mock_get_config(monkeypatch):
             _mock_length_from_prompt_token_ids_or_embeds,
         )
 
-    monkeypatch.setattr(
-        "vllm_omni.entrypoints.async_omni.init_tokenizer_from_configs", _mock_init_tokenizer_from_configs, raising=False
-    )
-    async_omni_path = "vllm_omni.entrypoints.async_omni"
-    if async_omni_path in sys.modules:
-        setattr(sys.modules[async_omni_path], "init_tokenizer_from_configs", _mock_init_tokenizer_from_configs)
-
     fake_hf_config = _ns()
     fake_hf_config.model_type = "qwen2_5_omni"