chore: import upstream snapshot with attribution

tests/entrypoints/multimodal/conftest.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from collections.abc import Callable, Iterator
+from contextlib import contextmanager
+from typing import Any
+
+import pytest
+
+# Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
+TEST_IMAGE_ASSETS = [
+    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
+    "Grayscale_8bits_palette_sample_image.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/Grayscale_8bits_palette_sample_image.png",
+    "1280px-Venn_diagram_rgb.svg.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/1280px-Venn_diagram_rgb.svg.png",
+    "RGBA_comp.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/RGBA_comp.png",
+]
+
+
+def _shutdown_llm(llm: Any, gpu_memory_utilization: float) -> None:
+    from vllm.distributed import cleanup_dist_env_and_memory
+    from vllm.platforms import current_platform
+
+    try:
+        shutdown_timeout = 60.0 if current_platform.is_rocm() else None
+        llm.llm_engine.engine_core.shutdown(timeout=shutdown_timeout)
+    except Exception:
+        pass
+
+    del llm
+
+    try:
+        import torch
+
+        torch._dynamo.reset()
+    except Exception:
+        pass
+
+    cleanup_dist_env_and_memory()
+
+    if current_platform.is_rocm():
+        from tests.utils import wait_for_rocm_memory_to_settle
+
+        wait_for_rocm_memory_to_settle(threshold_ratio=1.0 - gpu_memory_utilization)
+
+
+@contextmanager
+def managed_llm(*args: Any, **kwargs: Any) -> Iterator[Any]:
+    from vllm import LLM
+
+    llm = LLM(*args, **kwargs)
+    gpu_memory_utilization = (
+        llm.llm_engine.vllm_config.cache_config.gpu_memory_utilization
+    )
+    try:
+        yield llm
+    finally:
+        _shutdown_llm(llm, gpu_memory_utilization)
+
+
+def _make_managed_llm_factory() -> Iterator[Callable[..., Any]]:
+    from vllm import LLM
+
+    llms: list[tuple[Any, float]] = []
+
+    def make_llm(*args: Any, **kwargs: Any) -> Any:
+        llm = LLM(*args, **kwargs)
+        gpu_memory_utilization = (
+            llm.llm_engine.vllm_config.cache_config.gpu_memory_utilization
+        )
+        llms.append((llm, gpu_memory_utilization))
+        return llm
+
+    try:
+        yield make_llm
+    finally:
+        while llms:
+            llm, gpu_memory_utilization = llms.pop()
+            _shutdown_llm(llm, gpu_memory_utilization)
+
+
+@pytest.fixture
+def multimodal_llm_factory() -> Iterator[Callable[..., Any]]:
+    yield from _make_managed_llm_factory()

tests/entrypoints/multimodal/llm/test_chat.py

@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+
+from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
+
+
+@pytest.fixture(scope="function")
+def vision_llm(multimodal_llm_factory):
+    return multimodal_llm_factory(
+        model="microsoft/Phi-3.5-vision-instruct",
+        max_model_len=4096,
+        max_num_seqs=5,
+        enforce_eager=True,
+        trust_remote_code=True,
+        limit_mm_per_prompt={"image": 2},
+        seed=0,
+    )
+
+
+@pytest.mark.parametrize(
+    "image_urls", [[TEST_IMAGE_ASSETS[0], TEST_IMAGE_ASSETS[1]]], indirect=True
+)
+def test_chat_multi_image(vision_llm, image_urls: list[str]):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                *(
+                    {"type": "image_url", "image_url": {"url": image_url}}
+                    for image_url in image_urls
+                ),
+                {"type": "text", "text": "What's in this image?"},
+            ],
+        }
+    ]
+    outputs = vision_llm.chat(messages)
+    assert len(outputs) >= 0

tests/entrypoints/multimodal/llm/test_mm_cache_external_injection.py

@@ -0,0 +1,195 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Test that ``InputProcessor.inject_into_mm_cache()`` correctly injects
+pre-processed mm_kwargs into the processor cache and reports MM cache
+hit rate metrics accurately.
+
+This is used by frameworks like Dynamo that run the HF processor on a
+frontend and transfer pre-processed mm_kwargs to the backend, avoiding
+redundant processing.
+"""
+
+import logging
+
+import pytest
+import regex as re
+
+from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
+from vllm import LLM, SamplingParams
+from vllm.renderers.params import ChatParams
+from vllm.v1.metrics import loggers as stat_loggers
+from vllm.v1.metrics.reader import Counter, Metric
+
+
+def _make_messages(image_url: str):
+    return [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image_url",
+                    "image_url": {"url": image_url},
+                },
+            ],
+        }
+    ]
+
+
+def _get_counter_value(metrics: list[Metric], name: str):
+    metric = next(m for m in metrics if m.name == name)
+    assert isinstance(metric, Counter)
+    return metric.value
+
+
+def _get_mm_cache_stats(metrics: list[Metric]):
+    mm_cache_queries = _get_counter_value(metrics, "vllm:mm_cache_queries")
+    mm_cache_hits = _get_counter_value(metrics, "vllm:mm_cache_hits")
+    return mm_cache_queries, mm_cache_hits
+
+
+def _get_mm_cache_log(llm: LLM, caplog_vllm: pytest.LogCaptureFixture) -> float:
+    caplog_vllm.clear()
+    with caplog_vllm.at_level(logging.INFO, logger=stat_loggers.__name__):
+        llm.llm_engine.do_log_stats()
+
+    assert len(caplog_vllm.records) == 1
+    msg = caplog_vllm.records[0].getMessage()
+
+    assert "MM cache hit rate" in msg
+    match = re.search(r"MM cache hit rate: ([0-9.]+)%", msg)
+    assert match is not None
+    return float(match.group(1))
+
+
+@pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:2]], indirect=True)
+@pytest.mark.parametrize("mm_processor_cache_type", ["lru", "shm"])
+def test_inject_into_mm_cache(
+    num_gpus_available,
+    image_urls,
+    mm_processor_cache_type,
+    caplog_vllm,
+    multimodal_llm_factory,
+):
+    """Test that inject_into_mm_cache() injects pre-processed mm_kwargs into
+    the processor cache and MM cache hit metrics are updated correctly.
+
+    Steps:
+    1. Two normal requests (same image) -> cache miss then hit (baseline)
+    2. Extract cached kwargs, call inject_into_mm_cache with a new hash,
+       then generate with a pre-rendered input -> verifies injection works
+    """
+    llm = multimodal_llm_factory(
+        model="llava-hf/llava-1.5-7b-hf",
+        max_model_len=4096,
+        max_num_seqs=5,
+        enforce_eager=True,
+        disable_log_stats=False,
+        limit_mm_per_prompt={"image": 2},
+        mm_processor_cache_type=mm_processor_cache_type,
+    )
+
+    # Step 1: Normal requests to populate the cache
+    llm.chat(_make_messages(image_urls[0]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (1, 0)
+
+    llm.chat(_make_messages(image_urls[0]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (2, 1)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(50.0)
+
+    # Step 2: Use a second image to get valid expanded tokens and
+    # placeholder positions via the renderer.
+    llm.chat(_make_messages(image_urls[1]))
+    queries_before = _get_mm_cache_stats(llm.get_metrics())[0]  # 3
+
+    renderer = llm.llm_engine.renderer
+    cache = renderer.mm_processor_cache
+    assert cache is not None, "Processor cache should be enabled"
+
+    _, eng_prompts = renderer.render_chat(
+        [_make_messages(image_urls[1])],
+        ChatParams(),
+    )
+    eng_input = eng_prompts[0]
+
+    # Inject pre-processed mm_kwargs with a NEW hash via public API
+    new_mm_hash = "deadbeef" * 8
+    mm_hashes = {"image": [new_mm_hash]}
+    mm_kwargs = eng_input["mm_kwargs"]
+
+    llm.llm_engine.input_processor.inject_into_mm_cache(mm_hashes, mm_kwargs)
+
+    # Build pre-rendered input (no externally_processed flag needed)
+    pre_rendered_input = {
+        "type": "multimodal",
+        "prompt_token_ids": eng_input["prompt_token_ids"],
+        "mm_kwargs": mm_kwargs,
+        "mm_hashes": mm_hashes,
+        "mm_placeholders": eng_input["mm_placeholders"],
+    }
+
+    llm.generate(
+        pre_rendered_input,
+        sampling_params=SamplingParams(max_tokens=1),
+    )
+
+    # Verify cache was queried and injection happened
+    queries_after = _get_mm_cache_stats(llm.get_metrics())[0]
+    assert queries_after > queries_before, (
+        "Cache should have been queried for the injected item"
+    )
+    mm_rate = _get_mm_cache_log(llm, caplog_vllm)
+    assert mm_rate >= 0.0, "MM cache hit rate should be reported"
+
+
+@pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:1]], indirect=True)
+def test_inject_into_mm_cache_without_cache(
+    num_gpus_available,
+    image_urls,
+    multimodal_llm_factory,
+):
+    """Test that inject_into_mm_cache works gracefully when processor cache
+    is disabled (mm_processor_cache_gb=0). Should not crash.
+    """
+    llm = multimodal_llm_factory(
+        model="llava-hf/llava-1.5-7b-hf",
+        max_model_len=4096,
+        max_num_seqs=5,
+        enforce_eager=True,
+        disable_log_stats=False,
+        limit_mm_per_prompt={"image": 2},
+        mm_processor_cache_gb=0,
+    )
+
+    # Run a normal chat request first to warm up the model.
+    llm.chat(_make_messages(image_urls[0]))
+
+    # Use the renderer to get a proper EngineInput with expanded tokens
+    renderer = llm.llm_engine.renderer
+    _, eng_prompts = renderer.render_chat(
+        [_make_messages(image_urls[0])],
+        ChatParams(),
+    )
+    eng_input = eng_prompts[0]
+
+    mm_hashes = {"image": ["abcd1234" * 8]}
+    mm_kwargs = eng_input["mm_kwargs"]
+
+    # inject_into_mm_cache should not crash even without cache
+    llm.llm_engine.input_processor.inject_into_mm_cache(mm_hashes, mm_kwargs)
+
+    # Build and generate with pre-rendered input
+    pre_rendered_input = {
+        "type": "multimodal",
+        "prompt_token_ids": eng_input["prompt_token_ids"],
+        "mm_kwargs": mm_kwargs,
+        "mm_hashes": mm_hashes,
+        "mm_placeholders": eng_input["mm_placeholders"],
+    }
+
+    result = llm.generate(
+        pre_rendered_input,
+        sampling_params=SamplingParams(max_tokens=1),
+    )
+    assert len(result) == 1, "Should produce one output"
+    assert len(result[0].outputs) >= 1, "Should have at least one output sequence"

tests/entrypoints/multimodal/llm/test_mm_cache_stats.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import logging
+
+import pytest
+import regex as re
+
+from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
+from vllm import LLM
+from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
+from vllm.v1.metrics import loggers as stat_loggers
+from vllm.v1.metrics.reader import Counter, Metric
+
+
+def _make_messages(image_url: str) -> list[ChatCompletionMessageParam]:
+    return [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image_url",
+                    "image_url": {"url": image_url},
+                },
+            ],
+        }
+    ]
+
+
+def _get_counter_value(metrics: list[Metric], name: str):
+    metric = next(m for m in metrics if m.name == name)
+    assert isinstance(metric, Counter)
+    return metric.value
+
+
+def _get_mm_cache_stats(metrics: list[Metric]):
+    mm_cache_queries = _get_counter_value(metrics, "vllm:mm_cache_queries")
+    mm_cache_hits = _get_counter_value(metrics, "vllm:mm_cache_hits")
+
+    return mm_cache_queries, mm_cache_hits
+
+
+def _get_mm_cache_log(llm: LLM, caplog_vllm: pytest.LogCaptureFixture) -> float:
+    caplog_vllm.clear()
+    with caplog_vllm.at_level(logging.INFO, logger=stat_loggers.__name__):
+        llm.llm_engine.do_log_stats()
+
+    assert len(caplog_vllm.records) == 1
+    msg = caplog_vllm.records[0].getMessage()
+
+    assert "MM cache hit rate" in msg
+    match = re.search(r"MM cache hit rate: ([0-9.]+)%", msg)
+    assert match is not None
+    return float(match.group(1))
+
+
+@pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:2]], indirect=True)
+@pytest.mark.parametrize("mm_processor_cache_type", ["lru", "shm"])
+def test_mm_cache_stats(
+    num_gpus_available,
+    image_urls,
+    mm_processor_cache_type,
+    caplog_vllm,
+    multimodal_llm_factory,
+):
+    llm = multimodal_llm_factory(
+        model="llava-hf/llava-1.5-7b-hf",
+        max_model_len=4096,
+        max_num_seqs=5,
+        enforce_eager=True,
+        mm_processor_cache_type=mm_processor_cache_type,
+        disable_log_stats=False,
+        limit_mm_per_prompt={"image": 2},
+    )
+
+    llm.chat(_make_messages(image_urls[0]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (1, 0)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(0.0)
+
+    llm.chat(_make_messages(image_urls[1]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (2, 0)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(0.0)
+
+    llm.chat(_make_messages(image_urls[0]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (3, 1)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(33.3)
+
+    # NOTE: This only resets hit rate stats in CachingMetrics
+    # The raw queries and hits counts remain unaffected
+    llm.reset_mm_cache()
+
+    llm.chat(_make_messages(image_urls[0]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (4, 1)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(0.0)
+
+    llm.chat(_make_messages(image_urls[1]))
+    assert _get_mm_cache_stats(llm.get_metrics()) == (5, 1)
+    assert _get_mm_cache_log(llm, caplog_vllm) == pytest.approx(0.0)

tests/entrypoints/multimodal/llm/test_mm_embeds_only.py

@@ -0,0 +1,60 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+from tests.entrypoints.multimodal.conftest import managed_llm
+from vllm import LLM, SamplingParams
+from vllm.assets.image import ImageAsset
+
+MODEL = "llava-hf/llava-1.5-7b-hf"
+PROMPT = "USER: <image>\nDescribe this image briefly.\nASSISTANT:"
+TEXT_ONLY_PROMPT = "USER: What is 2 + 2?\nASSISTANT:"
+
+
+@pytest.fixture(scope="module")
+def llm():
+    """LLM with enable_mm_embeds=True and all modality limits zeroed out."""
+    with managed_llm(
+        model=MODEL,
+        max_model_len=2048,
+        enforce_eager=True,
+        gpu_memory_utilization=0.8,
+        enable_mm_embeds=True,
+        limit_mm_per_prompt={"image": 0},
+    ) as llm:
+        yield llm
+
+
+@pytest.mark.skip_global_cleanup
+def test_generate_with_embedding(llm: LLM):
+    """Pre-computed embedding produces tokens without hanging."""
+    embedding = ImageAsset("stop_sign").image_embeds
+    outputs = llm.generate(
+        {"prompt": PROMPT, "multi_modal_data": {"image": embedding}},
+        sampling_params=SamplingParams(max_tokens=32, temperature=0.0),
+    )
+    assert len(outputs) == 1
+    assert len(outputs[0].outputs[0].text) > 0
+
+
+@pytest.mark.skip_global_cleanup
+def test_raw_image_rejected(llm: LLM):
+    """Raw image input is still rejected when limit=0."""
+    raw_image = ImageAsset("stop_sign").pil_image
+    with pytest.raises(ValueError, match=r"At most 0 image\(s\)"):
+        llm.generate(
+            {"prompt": PROMPT, "multi_modal_data": {"image": raw_image}},
+            sampling_params=SamplingParams(max_tokens=16),
+        )
+
+
+@pytest.mark.skip_global_cleanup
+def test_text_only_prompt(llm: LLM):
+    """Text-only prompts still work under this config."""
+    outputs = llm.generate(
+        TEXT_ONLY_PROMPT,
+        sampling_params=SamplingParams(max_tokens=16, temperature=0.0),
+    )
+    assert len(outputs) == 1
+    assert len(outputs[0].outputs[0].text) > 0

tests/entrypoints/multimodal/llm/test_mm_processor_kwargs.py

@@ -0,0 +1,288 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from types import SimpleNamespace
+from unittest.mock import Mock
+
+import pytest
+
+from vllm import LLM, SamplingParams
+
+
+def _make_mock_llm() -> LLM:
+    llm = object.__new__(LLM)
+    llm.model_config = SimpleNamespace(
+        runner_type="generate", enable_prompt_embeds=False
+    )
+    return llm
+
+
+def test_generate_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"num_crops": 4}
+    sampling_params = SamplingParams(max_tokens=1)
+
+    llm._run_completion = Mock(return_value=["ok"])
+
+    outputs = llm.generate(
+        "prompt",
+        sampling_params=sampling_params,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == ["ok"]
+    assert llm._run_completion.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_enqueue_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"do_resize": False}
+    sampling_params = SamplingParams(max_tokens=1)
+
+    llm._add_completion_requests = Mock(return_value=["req-0"])
+
+    request_ids = llm.enqueue(
+        "prompt",
+        sampling_params=sampling_params,
+        use_tqdm=False,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert request_ids == ["req-0"]
+    assert llm._add_completion_requests.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_chat_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"do_pan_and_scan": True}
+    sampling_params = SamplingParams(max_tokens=1)
+    messages = [{"role": "user", "content": "hello"}]
+
+    llm._run_chat = Mock(return_value=["ok"])
+
+    outputs = llm.chat(
+        messages,
+        sampling_params=sampling_params,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == ["ok"]
+    assert llm._run_chat.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_enqueue_chat_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"do_pan_and_scan": True}
+    sampling_params = SamplingParams(max_tokens=1)
+    messages = [{"role": "user", "content": "hello"}]
+
+    llm._add_chat_requests = Mock(return_value=["req-0"])
+
+    request_ids = llm.enqueue_chat(
+        messages,
+        sampling_params=sampling_params,
+        use_tqdm=False,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert request_ids == ["req-0"]
+    assert llm._add_chat_requests.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_run_chat_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"num_crops": 8}
+    sampling_params = SamplingParams(max_tokens=1)
+    messages = [{"role": "user", "content": "hello"}]
+    sentinel_output = ["done"]
+
+    llm._add_chat_requests = Mock()
+    llm._run_engine = Mock(return_value=sentinel_output)
+
+    outputs = llm._run_chat(
+        messages=messages,
+        params=sampling_params,
+        output_type=object,
+        use_tqdm=False,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == sentinel_output
+    assert llm._add_chat_requests.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_run_completion_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"min_pixels": 4 * 28 * 28}
+    sampling_params = SamplingParams(max_tokens=1)
+    sentinel_output = ["done"]
+
+    llm._add_completion_requests = Mock()
+    llm._run_engine = Mock(return_value=sentinel_output)
+
+    outputs = llm._run_completion(
+        prompts=["prompt"],
+        params=sampling_params,
+        output_type=object,
+        use_tqdm=False,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == sentinel_output
+    assert llm._add_completion_requests.call_args.kwargs["mm_processor_kwargs"] == (
+        mm_processor_kwargs
+    )
+
+
+def test_add_completion_requests_forwards_mm_processor_kwargs() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"max_dynamic_patch": 4}
+    sampling_params = SamplingParams(max_tokens=1)
+
+    llm._params_to_seq = Mock(return_value=[sampling_params])
+    llm._lora_request_to_seq = Mock(return_value=[None])
+    llm._priority_to_seq = Mock(return_value=[0])
+    llm._preprocess_cmpl_one = Mock(return_value={"prompt_token_ids": [1]})
+
+    captured_prompts = []
+
+    def fake_render_and_add_requests(*, prompts, **_kwargs):
+        captured_prompts.extend(prompts)
+        return ["req-0"]
+
+    llm._render_and_add_requests = Mock(side_effect=fake_render_and_add_requests)
+
+    request_ids = llm._add_completion_requests(
+        prompts=["prompt"],
+        params=sampling_params,
+        use_tqdm=False,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert request_ids == ["req-0"]
+    llm._preprocess_cmpl_one.assert_called_once_with(
+        "prompt",
+        None,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+    assert captured_prompts == [{"prompt_token_ids": [1]}]
+
+
+def test_preprocess_cmpl_applies_mm_processor_kwargs_to_renderer(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"num_crops": 8}
+    prompt = {"prompt": "<image>", "multi_modal_data": {"image": object()}}
+
+    renderer = Mock()
+    renderer.default_cmpl_tok_params = Mock()
+    renderer.default_cmpl_tok_params.with_kwargs.return_value = "tok-params"
+    renderer.render_cmpl.return_value = ["engine-input"]
+    llm.renderer = renderer
+
+    monkeypatch.setattr(
+        "vllm.entrypoints.offline_utils.parse_model_prompt",
+        lambda _model_config, parsed_prompt: parsed_prompt,
+    )
+
+    outputs = llm._preprocess_cmpl(
+        [prompt],
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == ["engine-input"]
+    renderer.render_cmpl.assert_called_once_with(
+        [prompt],
+        "tok-params",
+        prompt_extras={"mm_processor_kwargs": mm_processor_kwargs},
+    )
+
+
+def test_preprocess_cmpl_keeps_prompt_mm_processor_kwargs_when_no_override(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    llm = _make_mock_llm()
+    prompt = {
+        "prompt": "<image>",
+        "multi_modal_data": {"image": object()},
+        "mm_processor_kwargs": {"num_crops": 2},
+    }
+
+    renderer = Mock()
+    renderer.default_cmpl_tok_params = Mock()
+    renderer.default_cmpl_tok_params.with_kwargs.return_value = "tok-params"
+    renderer.render_cmpl.return_value = ["engine-input"]
+    llm.renderer = renderer
+
+    monkeypatch.setattr(
+        "vllm.entrypoints.offline_utils.parse_model_prompt",
+        lambda _model_config, parsed_prompt: parsed_prompt,
+    )
+
+    outputs = llm._preprocess_cmpl([prompt])
+
+    assert outputs == ["engine-input"]
+    renderer.render_cmpl.assert_called_once_with(
+        [prompt],
+        "tok-params",
+        prompt_extras=None,
+    )
+
+
+def test_preprocess_chat_applies_mm_processor_kwargs_to_renderer() -> None:
+    llm = _make_mock_llm()
+    mm_processor_kwargs = {"num_crops": 8}
+    messages = [[{"role": "user", "content": "Describe this image."}]]
+
+    renderer = Mock()
+    renderer.tokenizer = object()
+    renderer.default_chat_tok_params = Mock()
+    renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params"
+    renderer.render_chat.return_value = (messages, ["engine-input"])
+    llm.renderer = renderer
+
+    outputs = llm._preprocess_chat(
+        messages,
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
+
+    assert outputs == ["engine-input"]
+    call_args = renderer.render_chat.call_args
+    assert call_args.args[0] == messages
+    assert call_args.args[1].mm_processor_kwargs == mm_processor_kwargs
+    assert call_args.args[2] == "tok-params"
+    assert call_args.kwargs["prompt_extras"] == {
+        "mm_processor_kwargs": mm_processor_kwargs
+    }
+
+
+def test_preprocess_chat_omits_mm_processor_kwargs_when_no_override() -> None:
+    llm = _make_mock_llm()
+    messages = [[{"role": "user", "content": "Describe this image."}]]
+
+    renderer = Mock()
+    renderer.tokenizer = object()
+    renderer.default_chat_tok_params = Mock()
+    renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params"
+    renderer.render_chat.return_value = (messages, ["engine-input"])
+    llm.renderer = renderer
+
+    outputs = llm._preprocess_chat(messages)
+
+    assert outputs == ["engine-input"]
+    call_args = renderer.render_chat.call_args
+    assert call_args.args[0] == messages
+    assert call_args.args[1].mm_processor_kwargs is None
+    assert call_args.args[2] == "tok-params"
+    assert call_args.kwargs["prompt_extras"] is None

tests/entrypoints/multimodal/openai/chat_completion/test_audio.py

@@ -0,0 +1,397 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+
+from tests.utils import RemoteOpenAIServer
+from vllm.assets.audio import AudioAsset
+from vllm.multimodal.utils import encode_audio_base64, encode_audio_url, fetch_audio
+
+MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
+TEST_AUDIO_URLS = [
+    AudioAsset("winning_call").url,
+    AudioAsset("mary_had_lamb").url,
+]
+MAXIMUM_AUDIOS = 2
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--dtype",
+        "float32",
+        "--max-model-len",
+        "2048",
+        "--max-num-seqs",
+        "5",
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--limit-mm-per-prompt",
+        json.dumps({"audio": MAXIMUM_AUDIOS}),
+    ]
+
+    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="session")
+def base64_encoded_audio() -> dict[str, str]:
+    return {
+        audio_url: encode_audio_base64(*fetch_audio(audio_url))
+        for audio_url in TEST_AUDIO_URLS
+    }
+
+
+@pytest.fixture(scope="session")
+def url_encoded_audio() -> dict[str, str]:
+    return {
+        audio_url: encode_audio_url(*fetch_audio(audio_url))
+        for audio_url in TEST_AUDIO_URLS
+    }
+
+
+def dummy_messages_from_audio_url(
+    audio_urls: str | list[str],
+    content_text: str = "What's happening in this audio?",
+):
+    if isinstance(audio_urls, str):
+        audio_urls = [audio_urls]
+
+    return [
+        {
+            "role": "user",
+            "content": [
+                *(
+                    {"type": "audio_url", "audio_url": {"url": audio_url}}
+                    for audio_url in audio_urls
+                ),
+                {"type": "text", "text": content_text},
+            ],
+        }
+    ]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])
+async def test_single_chat_session_audio(
+    client: openai.AsyncOpenAI, model_name: str, audio_url: str
+):
+    messages = dummy_messages_from_audio_url(audio_url)
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length"
+    assert chat_completion.usage == openai.types.CompletionUsage(
+        completion_tokens=10, prompt_tokens=202, total_tokens=212
+    )
+
+    message = choice.message
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 10
+    assert message.role == "assistant"
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+    )
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])
+async def test_error_on_invalid_audio_url_type(
+    client: openai.AsyncOpenAI, model_name: str, audio_url: str
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "audio_url", "audio_url": audio_url},
+                {"type": "text", "text": "What's happening in this audio?"},
+            ],
+        }
+    ]
+
+    # audio_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        _ = await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])
+async def test_single_chat_session_audio_base64encoded(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    audio_url: str,
+    url_encoded_audio: dict[str, str],
+):
+    messages = dummy_messages_from_audio_url(url_encoded_audio[audio_url])
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length"
+    assert chat_completion.usage == openai.types.CompletionUsage(
+        completion_tokens=10, prompt_tokens=202, total_tokens=212
+    )
+
+    message = choice.message
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 10
+    assert message.role == "assistant"
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+    )
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", [TEST_AUDIO_URLS[0]])
+async def test_single_chat_session_input_audio(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    audio_url: str,
+    base64_encoded_audio: dict[str, str],
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "input_audio",
+                    "input_audio": {
+                        "data": base64_encoded_audio[audio_url],
+                        "format": "wav",
+                    },
+                },
+                {"type": "text", "text": "What's happening in this audio?"},
+            ],
+        }
+    ]
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        logprobs=True,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length"
+    assert chat_completion.usage == openai.types.CompletionUsage(
+        completion_tokens=10, prompt_tokens=202, total_tokens=212
+    )
+
+    message = choice.message
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 10
+    assert message.role == "assistant"
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+    )
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", TEST_AUDIO_URLS)
+async def test_chat_streaming_audio(
+    client: openai.AsyncOpenAI, model_name: str, audio_url: str
+):
+    messages = dummy_messages_from_audio_url(
+        audio_url, "What's a short title for this audio?"
+    )
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=8,
+        temperature=0.0,
+    )
+    output = chat_completion.choices[0].message.content
+    stop_reason = chat_completion.choices[0].finish_reason
+
+    # test streaming
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=8,
+        temperature=0.0,
+        stream=True,
+    )
+    chunks: list[str] = []
+    finish_reason_count = 0
+    async for chunk in stream:
+        delta = chunk.choices[0].delta
+        if delta.role:
+            assert delta.role == "assistant"
+        if delta.content:
+            chunks.append(delta.content)
+        if chunk.choices[0].finish_reason is not None:
+            finish_reason_count += 1
+    # finish reason should only return in last block
+    assert finish_reason_count == 1
+    assert chunk.choices[0].finish_reason == stop_reason
+    assert delta.content
+    assert "".join(chunks) == output
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("audio_url", TEST_AUDIO_URLS)
+async def test_chat_streaming_input_audio(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    audio_url: str,
+    base64_encoded_audio: dict[str, str],
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "input_audio",
+                    "input_audio": {
+                        "data": base64_encoded_audio[audio_url],
+                        "format": "wav",
+                    },
+                },
+                {"type": "text", "text": "What's a short title for this audio?"},
+            ],
+        }
+    ]
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=8,
+        temperature=0.0,
+    )
+    output = chat_completion.choices[0].message.content
+    stop_reason = chat_completion.choices[0].finish_reason
+
+    # test streaming
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=8,
+        temperature=0.0,
+        stream=True,
+    )
+    chunks: list[str] = []
+    finish_reason_count = 0
+    async for chunk in stream:
+        delta = chunk.choices[0].delta
+        if delta.role:
+            assert delta.role == "assistant"
+        if delta.content:
+            chunks.append(delta.content)
+        if chunk.choices[0].finish_reason is not None:
+            finish_reason_count += 1
+    # finish reason should only return in last block
+    assert finish_reason_count == 1
+    assert chunk.choices[0].finish_reason == stop_reason
+    assert delta.content
+    assert "".join(chunks) == output
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "audio_urls", [TEST_AUDIO_URLS, TEST_AUDIO_URLS + [TEST_AUDIO_URLS[0]]]
+)
+async def test_multi_audio_input(
+    client: openai.AsyncOpenAI, model_name: str, audio_urls: list[str]
+):
+    messages = dummy_messages_from_audio_url(audio_urls)
+
+    if len(audio_urls) > MAXIMUM_AUDIOS:
+        with pytest.raises(openai.BadRequestError):  # test multi-audio input
+            await client.chat.completions.create(
+                model=model_name,
+                messages=messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+            )
+
+        # the server should still work afterwards
+        completion = await client.completions.create(
+            model=model_name,
+            prompt=[0, 0, 0, 0, 0],
+            max_tokens=5,
+            temperature=0.0,
+        )
+        completion = completion.choices[0].text
+        assert completion is not None and len(completion) >= 0
+    else:
+        chat_completion = await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+        message = chat_completion.choices[0].message
+        assert message.content is not None and len(message.content) >= 0

tests/entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py

@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai
+import pybase64 as base64
+import pytest
+import pytest_asyncio
+
+from tests.conftest import VideoTestAssets
+from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
+
+MODEL_NAME = "Qwen/Qwen2.5-Omni-3B"
+
+
+@pytest.fixture(scope="module")
+def server():
+    # Use module scope so the server is started once and shared across all
+    # tests in this file. Starting a new vLLM server per test on XPU can
+    # cause the second server startup to hang silently and exceed the
+    # wait-for-server timeout, resulting in RuntimeError.
+    args = [
+        "--max-model-len",
+        "16384",
+        "--enforce-eager",
+        "--limit-mm-per-prompt",
+        json.dumps({"audio": 3, "video": 3}),
+        *ROCM_EXTRA_ARGS,
+    ]
+
+    with RemoteOpenAIServer(
+        MODEL_NAME,
+        args,
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.mark.core_model
+@pytest.mark.asyncio
+async def test_online_audio_in_video(
+    client: openai.AsyncOpenAI, video_assets: VideoTestAssets
+):
+    """Test video input with `audio_in_video=True`"""
+
+    # we don't use video_urls above because they missed audio stream.
+    video_path = video_assets[0].video_path
+    with open(video_path, "rb") as f:
+        video_base64 = base64.b64encode(f.read()).decode("utf-8")
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What's in this video?"},
+                {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+                },
+            ],
+        }
+    ]
+
+    # multi-turn to test mm processor cache as well
+    for turn in range(2):
+        chat_completion = await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=messages,
+            max_tokens=8,
+            temperature=0.0,
+            extra_body={
+                "mm_processor_kwargs": {
+                    "use_audio_in_video": True,
+                }
+            },
+        )
+
+        assert len(chat_completion.choices) == 1
+        choice = chat_completion.choices[0]
+        print(
+            f"[DEBUG][single-video] turn={turn} "
+            f"finish_reason={choice.finish_reason!r} "
+            f"content={choice.message.content!r} "
+            f"usage={chat_completion.usage}"
+        )
+        assert choice.finish_reason == "length"
+
+
+@pytest.mark.core_model
+@pytest.mark.asyncio
+async def test_online_audio_in_video_multi_videos(
+    client: openai.AsyncOpenAI, video_assets: VideoTestAssets
+):
+    """Test multi-video input with `audio_in_video=True`"""
+
+    # we don't use video_urls above because they missed audio stream.
+    video_path = video_assets[0].video_path
+    with open(video_path, "rb") as f:
+        video_base64 = base64.b64encode(f.read()).decode("utf-8")
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What's in these two videos?"},
+                {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+                },
+                {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+                },
+            ],
+        }
+    ]
+
+    # multi-turn to test mm processor cache as well
+    for turn in range(2):
+        chat_completion = await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=messages,
+            max_tokens=8,
+            temperature=0.0,
+            extra_body={
+                "mm_processor_kwargs": {
+                    "use_audio_in_video": True,
+                }
+            },
+        )
+
+        assert len(chat_completion.choices) == 1
+        choice = chat_completion.choices[0]
+        print(
+            f"[DEBUG][multi-video] turn={turn} "
+            f"finish_reason={choice.finish_reason!r} "
+            f"content={choice.message.content!r} "
+            f"usage={chat_completion.usage}"
+        )
+        assert choice.finish_reason == "length"
+
+
+@pytest.mark.core_model
+@pytest.mark.asyncio
+async def test_online_audio_in_video_interleaved(
+    client: openai.AsyncOpenAI, video_assets: VideoTestAssets
+):
+    """Test interleaved video/audio input with `audio_in_video=True`"""
+
+    # we don't use video_urls above because they missed audio stream.
+    video_path = video_assets[0].video_path
+    with open(video_path, "rb") as f:
+        video_base64 = base64.b64encode(f.read()).decode("utf-8")
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What's in these two videos?"},
+                {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+                },
+                {
+                    "type": "audio_url",
+                    "audio_url": {"url": f"data:audio/mp4;base64,{video_base64}"},
+                },
+                {
+                    "type": "video_url",
+                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+                },
+            ],
+        }
+    ]
+    with pytest.raises(
+        openai.BadRequestError,
+        match="use_audio_in_video requires equal number of audio and video items",
+    ):
+        await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=messages,
+            max_tokens=16,
+            extra_body={
+                "mm_processor_kwargs": {
+                    "use_audio_in_video": True,
+                }
+            },
+        )

tests/entrypoints/multimodal/openai/chat_completion/test_chat_completion_with_image_embeds.py

@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai  # use the official client for correctness check
+import pytest
+import pytest_asyncio
+import torch
+from transformers import AutoConfig
+
+from tests.conftest import ImageTestAssets
+from tests.utils import RemoteOpenAIServer
+from vllm.utils.serial_utils import tensor2base64
+
+# any model with a chat template should work here
+MODEL_NAME = "llava-hf/llava-1.5-7b-hf"
+CONFIG = AutoConfig.from_pretrained(MODEL_NAME)
+MAXIMUM_IMAGES = 2
+
+
+@pytest.fixture(scope="module")
+def default_image_embeds_server_args() -> list[str]:
+    return [
+        "--dtype",
+        "bfloat16",
+        "--max-model-len",
+        "2048",
+        "--max-num-seqs",
+        "4",
+        "--enforce-eager",
+        "--limit-mm-per-prompt",
+        json.dumps({"image": MAXIMUM_IMAGES}),
+        "--enable-mm-embeds",
+    ]
+
+
+@pytest.fixture(scope="module")
+def server_with_image_embeds(default_image_embeds_server_args):
+    with RemoteOpenAIServer(
+        MODEL_NAME, default_image_embeds_server_args, max_wait_seconds=600
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client_with_image_embeds(server_with_image_embeds):
+    async with server_with_image_embeds.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("dtype", [torch.half, torch.float16, torch.float32])
+async def test_chat_completions_with_image_embeds(
+    client_with_image_embeds: openai.AsyncOpenAI,
+    model_name: str,
+    image_assets: ImageTestAssets,
+    dtype: torch.dtype,
+):
+    # Test case: Single image embeds input
+    image_embeds = image_assets[0].image_embeds.to(dtype=dtype)
+    base64_image_embedding = tensor2base64(image_embeds)
+    chat_completion = await client_with_image_embeds.chat.completions.create(
+        messages=[
+            {"role": "system", "content": "You are a helpful assistant."},
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": "Describe these images separately. For each image,"
+                        "reply with a short sentence (no more than 10 words).",
+                    },
+                    {
+                        "type": "image_embeds",
+                        "image_embeds": base64_image_embedding,
+                    },
+                ],
+            },
+        ],
+        model=model_name,
+    )
+    assert chat_completion.choices[0].message.content is not None
+    assert isinstance(chat_completion.choices[0].message.content, str)
+    assert len(chat_completion.choices[0].message.content) > 0

tests/entrypoints/multimodal/openai/chat_completion/test_chat_completion_with_mixed_audio_embeds.py

@@ -0,0 +1,190 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""E2E test for mixing `prompt_embeds` with `audio_embeds` in a single
+Chat Completions request."""
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+import safetensors
+import torch
+import torch.nn as nn
+from huggingface_hub import hf_hub_download
+from transformers import AutoConfig, AutoTokenizer
+
+from tests.utils import RemoteOpenAIServer
+from vllm.utils.serial_utils import tensor2base64
+
+QWEN2AUDIO_MODEL = "Qwen/Qwen2-Audio-7B-Instruct"
+
+# Use the model's native dtype to avoid an implicit cast inside
+# `safe_load_prompt_embeds` (mismatched floating-point dtypes are cast to the
+# model's dtype automatically, matching here just skips the conversion).
+QWEN2AUDIO_DTYPE = torch.bfloat16
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_server_args() -> list[str]:
+    return [
+        "--dtype",
+        "bfloat16",
+        "--max-model-len",
+        "2048",
+        "--max-num-seqs",
+        "4",
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--gpu-memory-utilization",
+        "0.85",
+        "--limit-mm-per-prompt",
+        json.dumps({"audio": 1}),
+        "--enable-prompt-embeds",
+        "--enable-mm-embeds",
+    ]
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_server(qwen2audio_server_args):
+    with RemoteOpenAIServer(
+        QWEN2AUDIO_MODEL,
+        qwen2audio_server_args,
+        max_wait_seconds=600,
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def qwen2audio_client(qwen2audio_server):
+    async with qwen2audio_server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_hidden_size() -> int:
+    config = AutoConfig.from_pretrained(QWEN2AUDIO_MODEL, trust_remote_code=True)
+    return config.text_config.hidden_size
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_prompt_embeds_b64(qwen2audio_hidden_size: int) -> str:
+    tensor = torch.randn(4, qwen2audio_hidden_size, dtype=QWEN2AUDIO_DTYPE)
+    return tensor2base64(tensor)
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_audio_embeds_b64(qwen2audio_hidden_size: int) -> str:
+    # Shape matches the `audio_embeds` unit-test fixture.
+    torch.manual_seed(0)
+    tensor = torch.randn(1, 128, qwen2audio_hidden_size, dtype=QWEN2AUDIO_DTYPE)
+    return tensor2base64(tensor)
+
+
+@pytest.mark.asyncio
+async def test_prompt_embeds_plus_audio_embeds(
+    qwen2audio_client: openai.AsyncOpenAI,
+    qwen2audio_prompt_embeds_b64: str,
+    qwen2audio_audio_embeds_b64: str,
+):
+    """Single user message carrying both prompt_embeds and audio_embeds parts."""
+    chat = await qwen2audio_client.chat.completions.create(
+        model=QWEN2AUDIO_MODEL,
+        max_tokens=5,
+        temperature=0.0,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "prompt_embeds",
+                        "data": qwen2audio_prompt_embeds_b64,
+                    },
+                    {
+                        "type": "audio_embeds",
+                        "audio_embeds": qwen2audio_audio_embeds_b64,
+                    },
+                    {"type": "text", "text": "Continue."},
+                ],
+            }
+        ],
+    )
+    assert chat.choices[0].message.content is not None
+    assert len(chat.choices[0].message.content) > 0
+
+
+@pytest.fixture(scope="module")
+def qwen2audio_aligned_content_and_embeds_b64() -> tuple[str, str]:
+    """Return `(content, base64_embeds)` where the embeddings are the model's
+    embedding of `content` tokenized WITHOUT special tokens.
+
+    Loads only the `embed_tokens` shard from disk on CPU (~1.1 GB of host
+    RAM) instead of the full 7B model on GPU.
+    """
+    content = "Describe this audio."
+    tokenizer = AutoTokenizer.from_pretrained(QWEN2AUDIO_MODEL, trust_remote_code=True)
+
+    index_path = hf_hub_download(QWEN2AUDIO_MODEL, "model.safetensors.index.json")
+    with open(index_path) as f:
+        weight_map = json.load(f)["weight_map"]
+    embed_key = next(k for k in weight_map if k.endswith("embed_tokens.weight"))
+    shard_path = hf_hub_download(QWEN2AUDIO_MODEL, weight_map[embed_key])
+    with safetensors.safe_open(shard_path, framework="pt", device="cpu") as f:
+        embed_weight = f.get_tensor(embed_key)
+    embed_layer = nn.Embedding.from_pretrained(embed_weight.to(QWEN2AUDIO_DTYPE))
+
+    ids = tokenizer(content, add_special_tokens=False, return_tensors="pt").input_ids
+    embeds = embed_layer(ids).squeeze(0)
+    return content, tensor2base64(embeds)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "audio_first",
+    [True, False],
+    ids=["audio_embeds-then-text", "text-then-audio_embeds"],
+)
+async def test_text_content_and_prompt_embeds_match_with_audio_embeds(
+    qwen2audio_client: openai.AsyncOpenAI,
+    qwen2audio_audio_embeds_b64: str,
+    qwen2audio_aligned_content_and_embeds_b64: tuple[str, str],
+    audio_first: bool,
+):
+    """Same content as text vs `prompt_embeds` should yield identical Chat
+    Completions output when mixed with `audio_embeds` in the same message.
+    """
+    content, encoded_text_embeds = qwen2audio_aligned_content_and_embeds_b64
+
+    audio_part = {
+        "type": "audio_embeds",
+        "audio_embeds": qwen2audio_audio_embeds_b64,
+    }
+    text_part = {"type": "text", "text": content}
+    embeds_part = {"type": "prompt_embeds", "data": encoded_text_embeds}
+
+    if audio_first:
+        text_content = [audio_part, text_part]
+        embeds_content = [audio_part, embeds_part]
+    else:
+        text_content = [text_part, audio_part]
+        embeds_content = [embeds_part, audio_part]
+
+    text_resp = await qwen2audio_client.chat.completions.create(
+        model=QWEN2AUDIO_MODEL,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": text_content}],
+    )
+    embeds_resp = await qwen2audio_client.chat.completions.create(
+        model=QWEN2AUDIO_MODEL,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": embeds_content}],
+    )
+
+    text_out = text_resp.choices[0].message.content
+    embeds_out = embeds_resp.choices[0].message.content
+    assert text_out is not None and len(text_out) > 0
+    assert embeds_out is not None and len(embeds_out) > 0
+    assert text_out == embeds_out

tests/entrypoints/multimodal/openai/chat_completion/test_chat_completion_with_mixed_image_embeds.py

@@ -0,0 +1,212 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""E2E tests for mixing `prompt_embeds` with image content parts in a single
+Chat Completions request.
+"""
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+import safetensors
+import torch
+import torch.nn as nn
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer
+
+from tests.utils import RemoteOpenAIServer
+from vllm.assets.image import ImageAsset
+from vllm.multimodal.utils import encode_image_url
+from vllm.utils.serial_utils import tensor2base64
+
+MODEL_NAME = "Qwen/Qwen2-VL-2B-Instruct"
+
+# Use the model's native dtype to skip the implicit cast inside
+# `safe_load_prompt_embeds` (mismatched floating-point dtypes are cast to the
+# model's dtype automatically).
+MODEL_DTYPE = torch.bfloat16
+
+
+@pytest.fixture(scope="module")
+def server_args() -> list[str]:
+    return [
+        "--dtype",
+        "bfloat16",
+        "--max-model-len",
+        "2048",
+        "--max-num-seqs",
+        "4",
+        "--enforce-eager",
+        "--gpu-memory-utilization",
+        "0.4",
+        "--limit-mm-per-prompt",
+        json.dumps({"image": 1}),
+        "--enable-prompt-embeds",
+        "--enable-mm-embeds",
+    ]
+
+
+@pytest.fixture(scope="module")
+def server(server_args):
+    with RemoteOpenAIServer(
+        MODEL_NAME,
+        server_args,
+        max_wait_seconds=600,
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="module")
+def image_url() -> str:
+    """Stable real image as a data URL, kept identical across both the
+    text and prompt_embeds requests so any output difference must come from
+    how the text content is delivered."""
+    return encode_image_url(ImageAsset("stop_sign").pil_image)
+
+
+@pytest.fixture(scope="module")
+def aligned_content_and_embeds_b64() -> tuple[str, str]:
+    """`(content, base64_embeds)` where the embeddings are the model's
+    embedding of `content` tokenized WITHOUT special tokens.
+
+    Loads only the `embed_tokens` shard from disk on CPU instead of the full
+    model on GPU, so the fixture has zero VRAM footprint and won't contend
+    with the running vLLM server.
+    """
+    content = "Describe this image."
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
+
+    index_path = hf_hub_download(MODEL_NAME, "model.safetensors.index.json")
+    with open(index_path) as f:
+        weight_map = json.load(f)["weight_map"]
+    embed_key = next(k for k in weight_map if k.endswith("embed_tokens.weight"))
+    shard_path = hf_hub_download(MODEL_NAME, weight_map[embed_key])
+    with safetensors.safe_open(shard_path, framework="pt", device="cpu") as f:
+        embed_weight = f.get_tensor(embed_key)
+    embed_layer = nn.Embedding.from_pretrained(embed_weight.to(MODEL_DTYPE))
+
+    ids = tokenizer(content, add_special_tokens=False, return_tensors="pt").input_ids
+    embeds = embed_layer(ids).squeeze(0)
+    return content, tensor2base64(embeds)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "image_first",
+    [True, False],
+    ids=["image_url-then-text", "text-then-image_url"],
+)
+async def test_text_content_and_prompt_embeds_match_with_image_url(
+    client: openai.AsyncOpenAI,
+    image_url: str,
+    aligned_content_and_embeds_b64: tuple[str, str],
+    image_first: bool,
+):
+    """Same content as text vs `prompt_embeds` should yield identical Chat
+    Completions output when mixed with an `image_url` part in the same
+    message under greedy decoding.
+    """
+    content, encoded_text_embeds = aligned_content_and_embeds_b64
+
+    image_part = {"type": "image_url", "image_url": {"url": image_url}}
+    text_part = {"type": "text", "text": content}
+    embeds_part = {"type": "prompt_embeds", "data": encoded_text_embeds}
+
+    if image_first:
+        text_content = [image_part, text_part]
+        embeds_content = [image_part, embeds_part]
+    else:
+        text_content = [text_part, image_part]
+        embeds_content = [embeds_part, image_part]
+
+    text_resp = await client.chat.completions.create(
+        model=MODEL_NAME,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": text_content}],
+    )
+    embeds_resp = await client.chat.completions.create(
+        model=MODEL_NAME,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": embeds_content}],
+    )
+
+    text_out = text_resp.choices[0].message.content
+    embeds_out = embeds_resp.choices[0].message.content
+    assert text_out is not None and len(text_out) > 0
+    assert embeds_out is not None and len(embeds_out) > 0
+    assert text_out == embeds_out
+
+
+@pytest.fixture(scope="module")
+def image_embeds_b64() -> dict[str, str]:
+    """Synthetic but stable `image_embeds` for Qwen2-VL."""
+    grid = (1, 4, 4)
+    spatial_merge_size = 2
+    num_patches = (grid[1] // spatial_merge_size) * (grid[2] // spatial_merge_size)
+    text_hidden_size = 1536  # Qwen2-VL-2B
+    torch.manual_seed(0)
+    return {
+        "image_embeds": tensor2base64(
+            torch.randn(num_patches, text_hidden_size, dtype=MODEL_DTYPE)
+        ),
+        "image_grid_thw": tensor2base64(torch.tensor(grid)),
+    }
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "image_first",
+    [True, False],
+    ids=["image_embeds-then-text", "text-then-image_embeds"],
+)
+async def test_text_content_and_prompt_embeds_match_with_image_embeds(
+    client: openai.AsyncOpenAI,
+    image_embeds_b64: dict[str, str],
+    aligned_content_and_embeds_b64: tuple[str, str],
+    image_first: bool,
+):
+    """Same content as text vs `prompt_embeds` should yield identical Chat
+    Completions output when mixed with a precomputed `image_embeds` part in
+    the same message under greedy decoding.
+    """
+    content, encoded_text_embeds = aligned_content_and_embeds_b64
+
+    image_part = {"type": "image_embeds", "image_embeds": image_embeds_b64}
+    text_part = {"type": "text", "text": content}
+    embeds_part = {"type": "prompt_embeds", "data": encoded_text_embeds}
+
+    if image_first:
+        text_content = [image_part, text_part]
+        embeds_content = [image_part, embeds_part]
+    else:
+        text_content = [text_part, image_part]
+        embeds_content = [embeds_part, image_part]
+
+    text_resp = await client.chat.completions.create(
+        model=MODEL_NAME,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": text_content}],
+    )
+    embeds_resp = await client.chat.completions.create(
+        model=MODEL_NAME,
+        max_tokens=10,
+        temperature=0.0,
+        messages=[{"role": "user", "content": embeds_content}],
+    )
+
+    text_out = text_resp.choices[0].message.content
+    embeds_out = embeds_resp.choices[0].message.content
+    assert text_out is not None and len(text_out) > 0
+    assert embeds_out is not None and len(embeds_out) > 0
+    assert text_out == embeds_out

tests/entrypoints/multimodal/openai/chat_completion/test_default_mm_loras.py

@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+
+import openai  # use the official client for correctness check
+import pytest
+import pytest_asyncio
+from huggingface_hub import snapshot_download
+
+from tests.conftest import AudioTestAssets
+from tests.utils import RemoteOpenAIServer
+
+# NOTE - the tests in this module are currently analogous to test_chat, but are
+# separated to avoid OOM killing due to module-scoped servers, since we
+# need a multimodal model for these tests.
+
+# Contains a modality specific lora alongside the base model
+MULTIMODAL_MODEL_NAME = snapshot_download("microsoft/Phi-4-multimodal-instruct")
+AUDIO_LORA_PATH = os.path.join(MULTIMODAL_MODEL_NAME, "speech-lora")
+
+ACTIVE_MM_LORA_RESPONSE = "Spoken text: The first words I spoke in the original chronograph, a little piece of practical poetry. Mary had a little lamb, it slept with quite a snow, and everywhere that Mary went, the lamb was sure to go."  # noqa: E501
+
+
+@pytest.fixture(scope="module")
+def multimodal_server():
+    args = [
+        # use half precision for speed and memory savings in CI environment
+        "--dtype",
+        "half",
+        "--max-model-len",
+        "4096",
+        "--enforce-eager",
+        # lora config below
+        "--enable-lora",
+        "--lora-modules",
+        f"speech={AUDIO_LORA_PATH}",
+        "--max-lora-rank",
+        "320",
+        "--max-num-seqs",
+        "2",
+        "--trust-remote-code",
+        "--gpu-memory-utilization",
+        "0.8",
+        "--default-mm-loras",
+        f'{{"audio": "{AUDIO_LORA_PATH}"}}',
+    ]
+
+    with RemoteOpenAIServer(
+        MULTIMODAL_MODEL_NAME, args, max_wait_seconds=480
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def multi_modal_client(multimodal_server):
+    async with multimodal_server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    # base model with default lora should give the same response as lora model
+    "model_name",
+    [MULTIMODAL_MODEL_NAME, "speech"],
+)
+async def test_default_mm_lora_chat_completions(
+    model_name: str,
+    multi_modal_client: openai.AsyncOpenAI,
+    audio_assets: AudioTestAssets,
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "text",
+                    "text": "Can you transcribe this audio?",
+                },
+                {
+                    "type": "audio_url",
+                    "audio_url": {"url": audio_assets[0].url},
+                },
+            ],
+        }
+    ]
+
+    chat_completion = await multi_modal_client.chat.completions.create(
+        model=model_name, messages=messages, max_completion_tokens=128, temperature=0.0
+    )
+
+    assert len(chat_completion.choices) > 0
+
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+    assert message.content == ACTIVE_MM_LORA_RESPONSE

tests/entrypoints/multimodal/openai/chat_completion/test_video.py

@@ -0,0 +1,403 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+
+from tests.utils import RemoteOpenAIServer
+from vllm.multimodal.utils import encode_video_url, fetch_video
+from vllm.platforms import current_platform
+
+MODEL_NAME = "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
+MAXIMUM_VIDEOS = 3
+
+TEST_VIDEO_URLS = [
+    "https://www.bogotobogo.com/python/OpenCV_Python/images/mean_shift_tracking/slow_traffic_small.mp4",
+    "https://github.com/opencv/opencv/raw/refs/tags/4.12.0/samples/data/vtest.avi",
+    "https://github.com/opencv/opencv/raw/refs/tags/4.12.0/samples/data/Megamind.avi",
+]
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--runner",
+        "generate",
+        "--max-model-len",
+        "32768",
+        "--max-num-seqs",
+        "2",
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--limit-mm-per-prompt",
+        json.dumps({"video": MAXIMUM_VIDEOS}),
+        "--media-io-kwargs",
+        json.dumps({"video": {"num_frames": 32}}),
+    ]
+
+    # ROCm: Increase timeouts to handle potential network delays and slower
+    # video processing when downloading multiple videos from external sources
+    env_overrides = {}
+    if current_platform.is_rocm():
+        env_overrides = {
+            "VLLM_VIDEO_FETCH_TIMEOUT": "120",
+            "VLLM_ENGINE_ITERATION_TIMEOUT_S": "300",
+        }
+
+    with RemoteOpenAIServer(MODEL_NAME, args, env_dict=env_overrides) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="session")
+def url_encoded_video() -> dict[str, str]:
+    return {
+        video_url: encode_video_url(fetch_video(video_url)[0])
+        for video_url in TEST_VIDEO_URLS
+    }
+
+
+def dummy_messages_from_video_url(
+    video_urls: str | list[str],
+    content_text: str = "What's in this video?",
+):
+    if isinstance(video_urls, str):
+        video_urls = [video_urls]
+
+    return [
+        {
+            "role": "user",
+            "content": [
+                *(
+                    {"type": "video_url", "video_url": {"url": video_url}}
+                    for video_url in video_urls
+                ),
+                {"type": "text", "text": content_text},
+            ],
+        }
+    ]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_single_chat_session_video(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = dummy_messages_from_video_url(video_url)
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length"
+    assert chat_completion.usage == openai.types.CompletionUsage(
+        completion_tokens=10, prompt_tokens=6287, total_tokens=6297
+    )
+
+    message = choice.message
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 10
+    assert message.role == "assistant"
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+    )
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", [TEST_VIDEO_URLS[0]])
+async def test_request_media_io_kwargs_override_uses_fewer_video_frames(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = dummy_messages_from_video_url(video_url)
+
+    default_resp = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=1,
+        temperature=0.0,
+    )
+    override_resp = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=1,
+        temperature=0.0,
+        extra_body={
+            "media_io_kwargs": {
+                "video": {
+                    "num_frames": 4,
+                }
+            }
+        },
+    )
+
+    assert default_resp.usage is not None
+    assert override_resp.usage is not None
+    assert override_resp.usage.prompt_tokens < default_resp.usage.prompt_tokens
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", [TEST_VIDEO_URLS[0]])
+async def test_invalid_num_frames_request_recoverable(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = dummy_messages_from_video_url(video_url)
+
+    with pytest.raises((openai.BadRequestError, openai.APIStatusError)):
+        await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=1,
+            temperature=0.0,
+            extra_body={
+                "media_io_kwargs": {
+                    "video": {
+                        "num_frames": "invalid",
+                    }
+                }
+            },
+        )
+
+    # Server should still handle subsequent requests after the failed one.
+    recovery_resp = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=1,
+        temperature=0.0,
+    )
+    recovery_msg = recovery_resp.choices[0].message
+    assert recovery_msg.content is not None and len(recovery_msg.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_error_on_invalid_video_url_type(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "video_url", "video_url": video_url},
+                {"type": "text", "text": "What's in this video?"},
+            ],
+        }
+    ]
+
+    # video_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        _ = await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_single_chat_session_video_beamsearch(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = dummy_messages_from_video_url(video_url)
+
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        n=2,
+        max_completion_tokens=10,
+        logprobs=True,
+        top_logprobs=5,
+        extra_body=dict(use_beam_search=True),
+    )
+    assert len(chat_completion.choices) == 2
+    assert (
+        chat_completion.choices[0].message.content
+        != chat_completion.choices[1].message.content
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_single_chat_session_video_base64encoded(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    video_url: str,
+    url_encoded_video: dict[str, str],
+):
+    messages = dummy_messages_from_video_url(url_encoded_video[video_url])
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length"
+    assert chat_completion.usage == openai.types.CompletionUsage(
+        completion_tokens=10, prompt_tokens=6287, total_tokens=6297
+    )
+
+    message = choice.message
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 10
+    assert message.role == "assistant"
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+    )
+    message = chat_completion.choices[0].message
+    assert message.content is not None and len(message.content) >= 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_single_chat_session_video_base64encoded_beamsearch(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    video_url: str,
+    url_encoded_video: dict[str, str],
+):
+    messages = dummy_messages_from_video_url(url_encoded_video[video_url])
+
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        n=2,
+        max_completion_tokens=10,
+        extra_body=dict(use_beam_search=True),
+    )
+    assert len(chat_completion.choices) == 2
+    assert (
+        chat_completion.choices[0].message.content
+        != chat_completion.choices[1].message.content
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+async def test_chat_streaming_video(
+    client: openai.AsyncOpenAI, model_name: str, video_url: str
+):
+    messages = dummy_messages_from_video_url(video_url)
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+    )
+    output = chat_completion.choices[0].message.content
+    stop_reason = chat_completion.choices[0].finish_reason
+
+    # test streaming
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+        stream=True,
+    )
+    chunks: list[str] = []
+    finish_reason_count = 0
+    async for chunk in stream:
+        delta = chunk.choices[0].delta
+        if delta.role:
+            assert delta.role == "assistant"
+        if delta.content:
+            chunks.append(delta.content)
+        if chunk.choices[0].finish_reason is not None:
+            finish_reason_count += 1
+    # finish reason should only return in last block
+    assert finish_reason_count == 1
+    assert chunk.choices[0].finish_reason == stop_reason
+    assert delta.content
+    assert "".join(chunks) == output
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "video_urls", [TEST_VIDEO_URLS[:i] for i in range(2, len(TEST_VIDEO_URLS))]
+)
+@pytest.mark.flaky(
+    reruns=2,
+    reruns_delay=5,
+    condition=current_platform.is_rocm(),
+)
+async def test_multi_video_input(
+    client: openai.AsyncOpenAI, model_name: str, video_urls: list[str]
+):
+    messages = dummy_messages_from_video_url(video_urls)
+
+    if len(video_urls) > MAXIMUM_VIDEOS:
+        with pytest.raises(openai.BadRequestError):  # test multi-video input
+            await client.chat.completions.create(
+                model=model_name,
+                messages=messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+            )
+
+        # the server should still work afterwards
+        completion = await client.completions.create(
+            model=model_name,
+            prompt=[0, 0, 0, 0, 0],
+            max_tokens=5,
+            temperature=0.0,
+        )
+        completion = completion.choices[0].text
+        assert completion is not None and len(completion) >= 0
+    else:
+        chat_completion = await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+        message = chat_completion.choices[0].message
+        assert message.content is not None and len(message.content) >= 0

tests/entrypoints/multimodal/openai/chat_completion/test_vision.py

@@ -0,0 +1,680 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+from transformers import AutoProcessor
+
+from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
+from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
+from vllm.multimodal.media import MediaWithBytes
+from vllm.multimodal.utils import encode_image_url, fetch_image
+from vllm.platforms import current_platform
+
+MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
+MAXIMUM_IMAGES = 2
+
+# Required terms for beam search validation
+# Each entry is a list of term groups - ALL groups must match
+# Each group is a list of alternatives - at least ONE term in the group must appear
+# This provides semantic validation while allowing wording variation
+REQUIRED_BEAM_SEARCH_TERMS = [
+    # Boardwalk image: must have "boardwalk" AND ("wooden" or "wood")
+    [["boardwalk"], ["wooden", "wood"]],
+    # Parrots image: must have ("parrot" or "bird") AND "two"
+    [["parrot", "bird"], ["two"]],
+    # Venn diagram: must have "venn" AND "diagram"
+    [["venn"], ["diagram"]],
+    # Gradient image: must have "gradient" AND ("color" or "spectrum")
+    [["gradient"], ["color", "spectrum"]],
+]
+
+
+def check_output_matches_terms(content: str, term_groups: list[list[str]]) -> bool:
+    """
+    Check if content matches all required term groups.
+    Each term group requires at least one of its terms to be present.
+    All term groups must be satisfied.
+    """
+    content_lower = content.lower()
+    return all(
+        any(term.lower() in content_lower for term in group) for group in term_groups
+    )
+
+
+def assert_non_empty_content(chat_completion, *, context: str = "") -> str:
+    """Assert the first choice has non-empty string content; return it.
+
+    Provides a detailed failure message including the full ChatCompletion
+    response so flaky / model-quality issues are easy to diagnose.
+    """
+    prefix = f"[{context}] " if context else ""
+    choice = chat_completion.choices[0]
+    content = choice.message.content
+
+    assert content is not None, (
+        f"{prefix}Expected non-None content but got None. "
+        f"finish_reason={choice.finish_reason!r}, "
+        f"full message={choice.message!r}, "
+        f"usage={chat_completion.usage!r}"
+    )
+    assert isinstance(content, str), (
+        f"{prefix}Expected str content, got {type(content).__name__}: {content!r}"
+    )
+    assert len(content) > 0, (
+        f"{prefix}Expected non-empty content but got empty string. "
+        f"finish_reason={choice.finish_reason!r}, "
+        f"full message={choice.message!r}, "
+        f"usage={chat_completion.usage!r}"
+    )
+    return content
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--runner",
+        "generate",
+        "--max-model-len",
+        "2048",
+        "--max-num-seqs",
+        "5",
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--limit-mm-per-prompt",
+        json.dumps({"image": MAXIMUM_IMAGES}),
+        *ROCM_EXTRA_ARGS,
+    ]
+
+    # ROCm: Increase timeouts to handle potential network delays and slower
+    # video processing when downloading multiple videos from external sources
+    env_overrides = {
+        **ROCM_ENV_OVERRIDES,
+        **(
+            {
+                "VLLM_VIDEO_FETCH_TIMEOUT": "120",
+                "VLLM_ENGINE_ITERATION_TIMEOUT_S": "300",
+            }
+            if current_platform.is_rocm()
+            else {}
+        ),
+    }
+
+    with RemoteOpenAIServer(MODEL_NAME, args, env_dict=env_overrides) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="session")
+def url_encoded_image(local_asset_server) -> dict[str, str]:
+    return {
+        image_asset: encode_image_url(local_asset_server.get_image_asset(image_asset))
+        for image_asset in TEST_IMAGE_ASSETS
+    }
+
+
+def dummy_messages_from_image_url(
+    image_urls: str | list[str],
+    content_text: str = "What's in this image?",
+):
+    if isinstance(image_urls, str):
+        image_urls = [image_urls]
+
+    return [
+        {
+            "role": "user",
+            "content": [
+                *(
+                    {"type": "image_url", "image_url": {"url": image_url}}
+                    for image_url in image_urls
+                ),
+                {"type": "text", "text": content_text},
+            ],
+        }
+    ]
+
+
+def describe_image_messages(
+    image_url: str, *, extra_image_fields: dict | None = None
+) -> list[dict]:
+    """Build the system + user messages used by the completions-with-image
+    family of tests. *extra_image_fields* is merged into the top-level
+    image content block (for uuid / bad-key tests)."""
+    image_block: dict = {
+        "type": "image_url",
+        "image_url": {"url": image_url},
+    }
+    if extra_image_fields:
+        image_block.update(extra_image_fields)
+
+    return [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Describe this image."},
+                image_block,
+            ],
+        },
+    ]
+
+
+async def complete_and_check(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    messages: list[dict],
+    *,
+    context: str,
+    max_completion_tokens: int = 50,
+    temperature: float = 0.0,
+) -> str:
+    """Run a chat completion and assert the output is non-empty.
+    Returns the content string."""
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=max_completion_tokens,
+        temperature=temperature,
+    )
+    return assert_non_empty_content(chat_completion, context=context)
+
+
+def get_hf_prompt_tokens(model_name, content, image_url):
+    processor = AutoProcessor.from_pretrained(
+        model_name, trust_remote_code=True, num_crops=4
+    )
+
+    placeholder = "<|image_1|>\n"
+    messages = [
+        {
+            "role": "user",
+            "content": f"{placeholder}{content}",
+        }
+    ]
+    image = fetch_image(image_url)
+    # Unwrap MediaWithBytes if present
+    if isinstance(image, MediaWithBytes):
+        image = image.media
+    images = [image]
+
+    prompt = processor.tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+    inputs = processor(prompt, images, return_tensors="pt")
+
+    return inputs.input_ids.shape[1]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_single_chat_session_image(
+    client: openai.AsyncOpenAI, model_name: str, image_url: str
+):
+    content_text = "What's in this image?"
+    messages = dummy_messages_from_image_url(image_url, content_text)
+
+    max_completion_tokens = 10
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=max_completion_tokens,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1, (
+        f"Expected 1 choice, got {len(chat_completion.choices)}"
+    )
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length", (
+        f"Expected finish_reason='length' (capped at {max_completion_tokens} "
+        f"tokens), got {choice.finish_reason!r}. "
+        f"content={choice.message.content!r}"
+    )
+
+    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text, image_url)
+    expected_usage = openai.types.CompletionUsage(
+        completion_tokens=max_completion_tokens,
+        prompt_tokens=hf_prompt_tokens,
+        total_tokens=hf_prompt_tokens + max_completion_tokens,
+    )
+    assert chat_completion.usage == expected_usage, (
+        f"Usage mismatch: got {chat_completion.usage!r}, expected {expected_usage!r}"
+    )
+
+    message = choice.message
+    assert message.content is not None and len(message.content) >= 10, (
+        f"Expected content with >=10 chars, got {message.content!r}"
+    )
+    assert message.role == "assistant", (
+        f"Expected role='assistant', got {message.role!r}"
+    )
+
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    await complete_and_check(
+        client,
+        model_name,
+        messages,
+        context=f"multi-turn follow-up for {image_url}",
+        max_completion_tokens=10,
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_error_on_invalid_image_url_type(
+    client: openai.AsyncOpenAI, model_name: str, image_url: str
+):
+    content_text = "What's in this image?"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": image_url},
+                {"type": "text", "text": content_text},
+            ],
+        }
+    ]
+
+    # image_url should be a dict {"url": "some url"}, not directly a string
+    with pytest.raises(openai.BadRequestError):
+        await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_single_chat_session_image_beamsearch(
+    client: openai.AsyncOpenAI, model_name: str, image_url: str
+):
+    content_text = "What's in this image?"
+    messages = dummy_messages_from_image_url(image_url, content_text)
+
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        n=2,
+        max_completion_tokens=10,
+        logprobs=True,
+        top_logprobs=5,
+        extra_body=dict(use_beam_search=True),
+    )
+    assert len(chat_completion.choices) == 2, (
+        f"Expected 2 beam search choices, got {len(chat_completion.choices)}"
+    )
+
+    content_0 = chat_completion.choices[0].message.content
+    content_1 = chat_completion.choices[1].message.content
+    assert content_0 != content_1, (
+        f"Beam search should produce different outputs for {image_url}, "
+        f"but both returned: {content_0!r}"
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_single_chat_session_image_base64encoded(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    raw_image_url: str,
+    image_url: str,
+    url_encoded_image: dict[str, str],
+):
+    content_text = "What's in this image?"
+    messages = dummy_messages_from_image_url(
+        url_encoded_image[raw_image_url],
+        content_text,
+    )
+
+    max_completion_tokens = 10
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=max_completion_tokens,
+        logprobs=True,
+        temperature=0.0,
+        top_logprobs=5,
+    )
+    assert len(chat_completion.choices) == 1, (
+        f"Expected 1 choice, got {len(chat_completion.choices)}"
+    )
+
+    choice = chat_completion.choices[0]
+    assert choice.finish_reason == "length", (
+        f"Expected finish_reason='length', got {choice.finish_reason!r}. "
+        f"content={choice.message.content!r}"
+    )
+
+    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text, image_url)
+    expected_usage = openai.types.CompletionUsage(
+        completion_tokens=max_completion_tokens,
+        prompt_tokens=hf_prompt_tokens,
+        total_tokens=hf_prompt_tokens + max_completion_tokens,
+    )
+    assert chat_completion.usage == expected_usage, (
+        f"Usage mismatch: got {chat_completion.usage!r}, expected {expected_usage!r}"
+    )
+
+    message = choice.message
+    assert message.content is not None and len(message.content) >= 10, (
+        f"Expected content with >=10 chars, got {message.content!r}"
+    )
+    assert message.role == "assistant", (
+        f"Expected role='assistant', got {message.role!r}"
+    )
+
+    messages.append({"role": "assistant", "content": message.content})
+
+    # test multi-turn dialogue
+    messages.append({"role": "user", "content": "express your result in json"})
+    await complete_and_check(
+        client,
+        model_name,
+        messages,
+        context=f"multi-turn base64 follow-up for {raw_image_url}",
+        max_completion_tokens=10,
+        temperature=0.0,
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_ASSETS))))
+async def test_single_chat_session_image_base64encoded_beamsearch(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    image_idx: int,
+    url_encoded_image: dict[str, str],
+):
+    # NOTE: This test validates that we pass MM data through beam search
+    raw_image_url = TEST_IMAGE_ASSETS[image_idx]
+    required_terms = REQUIRED_BEAM_SEARCH_TERMS[image_idx]
+
+    messages = dummy_messages_from_image_url(url_encoded_image[raw_image_url])
+
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        n=2,
+        max_completion_tokens=10,
+        temperature=0.0,
+        extra_body=dict(use_beam_search=True),
+    )
+    assert len(chat_completion.choices) == 2, (
+        f"Expected 2 beam search choices for image {image_idx} "
+        f"({raw_image_url}), got {len(chat_completion.choices)}"
+    )
+
+    # Verify beam search produces two different non-empty outputs
+    content_0 = chat_completion.choices[0].message.content
+    content_1 = chat_completion.choices[1].message.content
+
+    # Emit beam search outputs for debugging
+    print(
+        f"Beam search outputs for image {image_idx} ({raw_image_url}): "
+        f"Output 0: {content_0!r}, Output 1: {content_1!r}"
+    )
+
+    assert content_0, (
+        f"First beam output is empty for image {image_idx} ({raw_image_url}). "
+        f"finish_reason={chat_completion.choices[0].finish_reason!r}"
+    )
+    assert content_1, (
+        f"Second beam output is empty for image {image_idx} "
+        f"({raw_image_url}). "
+        f"finish_reason={chat_completion.choices[1].finish_reason!r}"
+    )
+    assert content_0 != content_1, (
+        f"Beam search produced identical outputs for image {image_idx} "
+        f"({raw_image_url}): {content_0!r}"
+    )
+
+    # Verify each output contains the required terms for this image
+    for i, content in enumerate([content_0, content_1]):
+        assert check_output_matches_terms(content, required_terms), (
+            f"Beam output {i} for image {image_idx} ({raw_image_url}) "
+            f"doesn't match required terms.\n"
+            f"  content: {content!r}\n"
+            f"  required (all groups, >=1 per group): {required_terms}"
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_chat_streaming_image(
+    client: openai.AsyncOpenAI, model_name: str, image_url: str
+):
+    messages = dummy_messages_from_image_url(image_url)
+
+    # test single completion
+    chat_completion = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+    )
+    output = chat_completion.choices[0].message.content
+    stop_reason = chat_completion.choices[0].finish_reason
+
+    # test streaming
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_completion_tokens=10,
+        temperature=0.0,
+        stream=True,
+    )
+    chunks: list[str] = []
+    finish_reason_count = 0
+    async for chunk in stream:
+        delta = chunk.choices[0].delta
+        if delta.role:
+            assert delta.role == "assistant", (
+                f"Expected role='assistant' in stream delta, got {delta.role!r}"
+            )
+        if delta.content:
+            chunks.append(delta.content)
+        if chunk.choices[0].finish_reason is not None:
+            finish_reason_count += 1
+    # finish reason should only return in last block
+    assert finish_reason_count == 1, (
+        f"Expected exactly 1 finish_reason across stream chunks, "
+        f"got {finish_reason_count}"
+    )
+    assert chunk.choices[0].finish_reason == stop_reason, (
+        f"Stream finish_reason={chunk.choices[0].finish_reason!r} "
+        f"doesn't match non-stream finish_reason={stop_reason!r}"
+    )
+
+    streamed_text = "".join(chunks)
+    assert streamed_text == output, (
+        f"Streamed output doesn't match non-streamed for {image_url}.\n"
+        f"  streamed:     {streamed_text!r}\n"
+        f"  non-streamed: {output!r}"
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "image_urls",
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True,
+)
+async def test_multi_image_input(
+    client: openai.AsyncOpenAI, model_name: str, image_urls: list[str]
+):
+    messages = dummy_messages_from_image_url(image_urls)
+
+    if len(image_urls) > MAXIMUM_IMAGES:
+        with pytest.raises(openai.BadRequestError):  # test multi-image input
+            await client.chat.completions.create(
+                model=model_name,
+                messages=messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+            )
+
+        # the server should still work afterwards
+        completion = await client.completions.create(
+            model=model_name,
+            prompt=[0, 0, 0, 0, 0],
+            max_tokens=5,
+            temperature=0.0,
+        )
+        assert completion.choices[0].text is not None, (
+            "Server failed to produce output after rejecting over-limit "
+            "multi-image request"
+        )
+    else:
+        await complete_and_check(
+            client,
+            model_name,
+            messages,
+            context=f"multi-image input ({len(image_urls)} images)",
+            max_completion_tokens=10,
+            temperature=0.0,
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "image_urls",
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True,
+)
+async def test_completions_with_image(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    image_urls: list[str],
+):
+    for image_url in image_urls:
+        messages = describe_image_messages(image_url)
+        await complete_and_check(
+            client,
+            model_name,
+            messages,
+            context=f"completions_with_image url={image_url}",
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "image_urls",
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True,
+)
+async def test_completions_with_image_with_uuid(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    image_urls: list[str],
+):
+    for image_url in image_urls:
+        messages = describe_image_messages(
+            image_url,
+            extra_image_fields={"uuid": image_url},
+        )
+        await complete_and_check(
+            client,
+            model_name,
+            messages,
+            context=f"uuid first request url={image_url}",
+        )
+
+        cached_messages: list[dict] = [
+            {"role": "system", "content": "You are a helpful assistant."},
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Describe this image."},
+                    {"type": "image_url", "image_url": {}, "uuid": image_url},
+                ],
+            },
+        ]
+        await complete_and_check(
+            client,
+            model_name,
+            cached_messages,
+            context=f"uuid cached (empty image) uuid={image_url}",
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_completions_with_empty_image_with_uuid_without_cache_hit(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+):
+    with pytest.raises(openai.BadRequestError):
+        await client.chat.completions.create(
+            messages=[
+                {"role": "system", "content": "You are a helpful assistant."},
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "Describe this image."},
+                        {
+                            "type": "image_url",
+                            "image_url": {},
+                            "uuid": "uuid_not_previously_seen",
+                        },
+                    ],
+                },
+            ],
+            model=model_name,
+        )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "image_urls",
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True,
+)
+async def test_completions_with_image_with_incorrect_uuid_format(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    image_urls: list[str],
+):
+    for image_url in image_urls:
+        messages = describe_image_messages(
+            image_url,
+            extra_image_fields={
+                "also_incorrect_uuid_key": image_url,
+            },
+        )
+        # Inject the bad key inside image_url dict too
+        messages[1]["content"][1]["image_url"]["incorrect_uuid_key"] = image_url
+
+        await complete_and_check(
+            client,
+            model_name,
+            messages,
+            context=f"incorrect uuid format url={image_url}",
+        )

tests/entrypoints/multimodal/openai/chat_completion/test_vision_embeds.py

@@ -0,0 +1,160 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import importlib.util
+
+import numpy as np
+import pybase64 as base64
+import pytest
+import requests
+import torch
+
+from tests.utils import RemoteOpenAIServer
+from vllm.utils.serial_utils import tensor2base64
+
+# Prithvi requires terratorch, which is temporarily unavailable while PyPI has
+# `lightning` quarantined (#41376). Skip just the Prithvi case; leave the
+# Qwen3-VL case in the same file untouched.
+_TERRATORCH_AVAILABLE = importlib.util.find_spec("terratorch") is not None
+
+
+@pytest.mark.skipif(
+    not _TERRATORCH_AVAILABLE,
+    reason="terratorch unavailable while PyPI has `lightning` quarantined; see #41376",
+)
+@pytest.mark.parametrize(
+    "model_name", ["ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11"]
+)
+def test_single_content(model_name: str):
+    args = [
+        "--runner",
+        "pooling",
+        # use half precision for speed and memory savings in CI environment
+        "--dtype",
+        "float16",
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--max-num-seqs",
+        "32",
+        "--model-impl",
+        "terratorch",
+        "--skip-tokenizer-init",
+        "--enable-mm-embeds",
+    ]
+
+    with RemoteOpenAIServer(model_name, args) as server:
+        response = requests.post(
+            server.url_for("pooling"),
+            json={
+                "model": model_name,
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "image_embeds",
+                                "image_embeds": {
+                                    "pixel_values": tensor2base64(
+                                        torch.ones((6, 512, 512), dtype=torch.float16)
+                                    ),
+                                    "location_coords": tensor2base64(
+                                        torch.ones((1, 2), dtype=torch.float16)
+                                    ),
+                                },
+                            },
+                        ],
+                    }
+                ],
+                "encoding_format": "base64",
+            },
+        )
+        response.raise_for_status()
+
+        output = response.json()["data"][0]["data"]
+
+        np_response = np.frombuffer(base64.b64decode(output), dtype=np.float32)
+        assert len(np_response) == 524288
+
+
+@pytest.mark.parametrize("model_name", ["Qwen/Qwen3-VL-2B-Instruct"])
+def test_multi_content(model_name: str):
+    args = [
+        "--enforce-eager",
+        "--max-num-seqs",
+        "32",
+        "--max-model-len",
+        "8192",
+        "--enable-mm-embeds",
+    ]
+
+    with RemoteOpenAIServer(model_name, args) as server:
+        client = server.get_client()
+
+        # Image only
+        chat_completion = client.chat.completions.create(
+            model=model_name,
+            messages=[
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_embeds",
+                            "image_embeds": {
+                                "image_embeds": tensor2base64(torch.zeros(220, 8192)),
+                                "image_grid_thw": tensor2base64(
+                                    torch.tensor([1, 22, 40])
+                                ),
+                            },
+                        },
+                        {
+                            "type": "image_embeds",
+                            "image_embeds": {
+                                "image_embeds": tensor2base64(torch.zeros(220, 8192)),
+                                "image_grid_thw": tensor2base64(
+                                    torch.tensor([1, 22, 40])
+                                ),
+                            },
+                        },
+                    ],
+                }
+            ],
+            max_tokens=5,
+        )
+
+        assert chat_completion.id is not None
+        assert len(chat_completion.choices) == 1
+
+        # Interleaved text and image
+        chat_completion = client.chat.completions.create(
+            model=model_name,
+            messages=[
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_embeds",
+                            "image_embeds": {
+                                "image_embeds": tensor2base64(torch.zeros(220, 8192)),
+                                "image_grid_thw": tensor2base64(
+                                    torch.tensor([1, 22, 40])
+                                ),
+                            },
+                        },
+                        {"type": "text", "text": "OCR:"},
+                        {
+                            "type": "image_embeds",
+                            "image_embeds": {
+                                "image_embeds": tensor2base64(torch.zeros(220, 8192)),
+                                "image_grid_thw": tensor2base64(
+                                    torch.tensor([1, 22, 40])
+                                ),
+                            },
+                        },
+                    ],
+                }
+            ],
+            max_tokens=5,
+        )
+
+        assert chat_completion.id is not None
+        assert len(chat_completion.choices) == 1

tests/entrypoints/multimodal/openai/responses/test_image.py

@@ -0,0 +1,165 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+
+import openai
+import pytest
+import pytest_asyncio
+
+from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
+from tests.utils import RemoteOpenAIServer
+from vllm.multimodal.utils import encode_image_url
+
+# Use a small vision model for testing
+MODEL_NAME = "Qwen/Qwen2.5-VL-3B-Instruct"
+MAXIMUM_IMAGES = 2
+
+
+@pytest.fixture(scope="module")
+def default_image_server_args():
+    return [
+        "--enforce-eager",
+        "--max-model-len",
+        "6000",
+        "--max-num-seqs",
+        "128",
+        "--limit-mm-per-prompt",
+        json.dumps({"image": MAXIMUM_IMAGES}),
+    ]
+
+
+@pytest.fixture(scope="module")
+def image_server(default_image_server_args):
+    with RemoteOpenAIServer(
+        MODEL_NAME,
+        default_image_server_args,
+        env_dict={"VLLM_ENABLE_RESPONSES_API_STORE": "1"},
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(image_server):
+    async with image_server.get_async_client() as async_client:
+        yield async_client
+
+
+@pytest.fixture(scope="session")
+def url_encoded_image(local_asset_server) -> dict[str, str]:
+    return {
+        image_url: encode_image_url(local_asset_server.get_image_asset(image_url))
+        for image_url in TEST_IMAGE_ASSETS
+    }
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
+async def test_single_chat_session_image(
+    client: openai.AsyncOpenAI, model_name: str, image_url: str
+):
+    content_text = "What's in this image?"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "input_image",
+                    "image_url": image_url,
+                    "detail": "auto",
+                },
+                {"type": "input_text", "text": content_text},
+            ],
+        }
+    ]
+
+    # test image url
+    response = await client.responses.create(
+        model=model_name,
+        input=messages,
+    )
+    assert len(response.output_text) > 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
+async def test_single_chat_session_image_base64encoded(
+    client: openai.AsyncOpenAI,
+    model_name: str,
+    raw_image_url: str,
+    url_encoded_image: dict[str, str],
+):
+    content_text = "What's in this image?"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "input_image",
+                    "image_url": url_encoded_image[raw_image_url],
+                    "detail": "auto",
+                },
+                {"type": "input_text", "text": content_text},
+            ],
+        }
+    ]
+    # test image base64
+    response = await client.responses.create(
+        model=model_name,
+        input=messages,
+    )
+    assert len(response.output_text) > 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+@pytest.mark.parametrize(
+    "image_urls",
+    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
+    indirect=True,
+)
+async def test_multi_image_input(
+    client: openai.AsyncOpenAI, model_name: str, image_urls: list[str]
+):
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                *(
+                    {
+                        "type": "input_image",
+                        "image_url": image_url,
+                        "detail": "auto",
+                    }
+                    for image_url in image_urls
+                ),
+                {"type": "input_text", "text": "What's in this image?"},
+            ],
+        }
+    ]
+
+    if len(image_urls) > MAXIMUM_IMAGES:
+        with pytest.raises(openai.BadRequestError):  # test multi-image input
+            await client.responses.create(
+                model=model_name,
+                input=messages,
+            )
+        # the server should still work afterwards
+        response = await client.responses.create(
+            model=model_name,
+            input=[
+                {
+                    "role": "user",
+                    "content": "What's the weather like in Paris today?",
+                }
+            ],
+        )
+        assert len(response.output_text) > 0
+    else:
+        response = await client.responses.create(
+            model=model_name,
+            input=messages,
+        )
+        assert len(response.output_text) > 0