[Rebase] Rebase to vllm v0.19.0 (#2475)

tests/diffusion/quantization/test_quantization_quality.py

@@ -118,8 +118,12 @@ def _generate_image(omni, config: QualityTestConfig):
 
     peak_mem = torch.cuda.max_memory_allocated() / (1024**3)
     first = outputs[0]
-    req_out = first.request_output[0] if hasattr(first, "request_output") else first
-    return req_out.images[0], peak_mem
+    if hasattr(first, "images") and first.images:
+        return first.images[0], peak_mem
+    inner = first.request_output
+    if inner is not None and hasattr(inner, "images") and inner.images:
+        return inner.images[0], peak_mem
+    raise ValueError("Could not extract image from output.")
 
 
 def _generate_video(omni, config: QualityTestConfig):

tests/e2e/offline_inference/test_bagel_text2img.py

@@ -229,6 +229,24 @@ def _wait_for_port(host: str, port: int, timeout: int = 30) -> bool:
     return False
 
 
+def _is_mooncake_master_available() -> bool:
+    """Check if mooncake_master binary is present and can actually execute."""
+    import shutil
+
+    binary = shutil.which("mooncake_master")
+    if binary is None:
+        return False
+    try:
+        result = subprocess.run(
+            [binary, "--help"],
+            capture_output=True,
+            timeout=5,
+        )
+        return result.returncode != 127
+    except (subprocess.TimeoutExpired, OSError):
+        return True
+
+
 def _cleanup_mooncake_processes(timeout_secs: int = 5) -> None:
     """Clean up any existing mooncake_master processes.
 
@@ -292,6 +310,8 @@ def _load_mooncake_config(host: str, rpc_port: int, http_port: int) -> str:
 @hardware_test(res={"cuda": "H100"})
 def test_bagel_text2img_mooncake_connector(run_level):
     """Test Bagel text2img with Mooncake connector for inter-stage communication."""
+    if not _is_mooncake_master_available():
+        pytest.skip("mooncake_master is not available or cannot execute (missing shared libraries like libibverbs)")
     MOONCAKE_HOST = "127.0.0.1"
     MOONCAKE_RPC_PORT = _find_free_port()
     MOONCAKE_HTTP_PORT = _find_free_port()

tests/e2e/online_serving/test_mimo_audio.py

@@ -63,20 +63,27 @@ def download_tokenizer():
 
 
 # CI stage config for H100 / MI325
-stage_configs = [get_chunk_config()]
-tokenizer_path = download_tokenizer()
-os.environ["MIMO_AUDIO_TOKENIZER_PATH"] = tokenizer_path
+# Guard module-level setup so test collection doesn't fail in environments
+# where the model cache is read-only or models aren't available.
+try:
+    stage_configs = [get_chunk_config()]
+    tokenizer_path = download_tokenizer()
+    os.environ["MIMO_AUDIO_TOKENIZER_PATH"] = tokenizer_path
 
-# Create parameter combinations for model and stage config
-test_params = [
-    OmniServerParams(
-        model=model,
-        stage_config_path=stage_config,
-        server_args=["--chat-template", CHAT_TEMPLATE_PATH],
+    test_params = [
+        OmniServerParams(
+            model=model,
+            stage_config_path=stage_config,
+            server_args=["--chat-template", CHAT_TEMPLATE_PATH],
+        )
+        for model in models
+        for stage_config in stage_configs
+    ]
+except Exception as exc:
+    pytest.skip(
+        f"MiMo-Audio online serving tests skipped: module setup failed ({type(exc).__name__}: {exc})",
+        allow_module_level=True,
     )
-    for model in models
-    for stage_config in stage_configs
-]
 
 
 def get_prompt(prompt_type="text_only"):

tests/engine/test_async_omni_engine_abort.py

@@ -60,6 +60,7 @@ async def generate(
 
 @pytest.mark.core_model
 @pytest.mark.omni
+@pytest.mark.real_hf_config
 @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
 @pytest.mark.asyncio
 async def test_abort():

tests/entrypoints/openai_api/test_serving_speech.py

@@ -1664,6 +1664,7 @@ class TestFishSpeechServing:
         assert all(allowed_special is None for _, _, allowed_special in tokenizer.calls)
 
     def test_build_fish_clone_prompt_normalizes_text_fields(self, fish_speech_server):
+        fish_speech_server._fish_speech_tokenizer = _FakeFishTokenizer()
         fish_speech_server._estimate_fish_prompt_len = MagicMock(return_value=123)
 
         request = OpenAICreateSpeechRequest(

tests/entrypoints/test_async_omni_abort.py

@@ -13,8 +13,8 @@ def test_generate_accepts_request_after_repeated_cancellations():
         submitted_request_ids = []
         aborted_request_batches = []
 
-        async def fake_add_request_async(*, request_id, prompt, sampling_params_list, final_stage_id):
-            del prompt, sampling_params_list, final_stage_id
+        async def fake_add_request_async(*, request_id, prompt, sampling_params_list, final_stage_id, **kwargs):
+            del prompt, sampling_params_list, final_stage_id, kwargs
             submitted_request_ids.append(request_id)
 
         async def fake_abort_async(request_ids):

tests/entrypoints/test_omni_entrypoints.py

@@ -113,6 +113,7 @@ class FakeAsyncOmniEngine:
         sampling_params_list: list[Any] | None = None,
         final_stage_id: int = 0,
         arrival_time: float | None = None,
+        **kwargs: Any,
     ) -> None:
         msg = {
             "request_id": request_id,

tests/examples/conftest.py

@@ -14,7 +14,6 @@ from collections.abc import Callable
 from pathlib import Path
 from typing import Any, NamedTuple, cast
 
-import mistune
 import pytest
 import torch
 from safetensors.torch import save_file
@@ -62,6 +61,8 @@ class ReadmeSnippet(NamedTuple):
         readme_path: Path,
         skipif: ReadmeSnippetExtractionSkipPredicate | None = None,
     ) -> list["ReadmeSnippet"]:
+        import mistune
+
         markdown = mistune.create_markdown(renderer="ast")
         tokens = markdown(readme_path.read_text(encoding="utf-8"))
         tokens = cast(list[dict[str, Any]], tokens)  # mistune's AST renderer always produces a list, not a str

tests/model_executor/models/test_omni_processing.py

@@ -21,9 +21,10 @@ from vllm.config.multimodal import (
     ImageDummyOptions,
     VideoDummyOptions,
 )
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
+from vllm.inputs import MultiModalDataDict, MultiModalInput
+from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.cache import MultiModalProcessorOnlyCache
-from vllm.multimodal.inputs import MultiModalInputs, batched_tensors_equal
+from vllm.multimodal.inputs import batched_tensors_equal
 from vllm.multimodal.processing import BaseMultiModalProcessor, InputProcessingContext
 from vllm.tokenizers import TokenizerLike, cached_tokenizer_from_config
 
@@ -302,8 +303,8 @@ def test_omni_processing_correctness(
 
 
 def _assert_inputs_equal(
-    a: MultiModalInputs,
-    b: MultiModalInputs,
+    a: MultiModalInput,
+    b: MultiModalInput,
     *,
     ignore_mm_keys: set[str] | None = None,
     msg: str = "",