Add vLLM v0.18.1 source tree with KV transfer abort fix

third_party/vllm/ now tracked in git for direct patch management.
Based on vLLM v0.18.1 release with one patch applied:

  vllm/v1/core/sched/scheduler.py:
    Replace fatal assert with graceful skip when KV transfer callback
    arrives for an already-aborted request during PD disaggregated serving.

Future vLLM modifications should be made directly in third_party/vllm/
and committed normally. The patches/ directory is kept as documentation
of what changed from upstream.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

third_party/vllm/tests/entrypoints/openai/cpu/test_render.py

@@ -0,0 +1,265 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Tests for the /render endpoints that expose prompt preprocessing."""
+
+import httpx
+import pytest
+import pytest_asyncio
+
+from tests.utils import RemoteLaunchRenderServer
+
+MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
+
+
+@pytest.fixture(scope="module")
+def server():
+    args: list[str] = []
+
+    with RemoteLaunchRenderServer(MODEL_NAME, args) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def client(server):
+    async with httpx.AsyncClient(
+        base_url=server.url_for(""), timeout=30.0
+    ) as http_client:
+        yield http_client
+
+
+@pytest.mark.asyncio
+async def test_completion_render_basic(client):
+    """Test basic completion render endpoint."""
+    # Make request to render endpoint
+    response = await client.post(
+        "/v1/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "prompt": "When should a chat-completions handler return an empty string?",
+        },
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+
+    # Verify response structure - list of GenerateRequest
+    assert isinstance(data, list)
+    assert len(data) > 0
+
+    # Verify first prompt is a GenerateRequest
+    first_prompt = data[0]
+    assert "token_ids" in first_prompt
+    assert "sampling_params" in first_prompt
+    assert "model" in first_prompt
+    assert "request_id" in first_prompt
+    assert isinstance(first_prompt["token_ids"], list)
+    assert len(first_prompt["token_ids"]) > 0
+    assert first_prompt["model"] == MODEL_NAME
+    assert first_prompt["request_id"].startswith("cmpl-")
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_basic(client):
+    """Test basic chat completion render endpoint."""
+    # Make request to render endpoint
+    response = await client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": (
+                        "Returning an empty string for the prompt may be confusing."
+                    ),
+                }
+            ],
+        },
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+
+    # Verify response structure - should be a GenerateRequest
+    assert isinstance(data, dict)
+    assert "token_ids" in data
+    assert isinstance(data["token_ids"], list)
+    assert len(data["token_ids"]) > 0
+
+    # Verify token IDs are integers and BOS token is present
+    token_ids = data["token_ids"]
+    assert all(isinstance(tid, int) for tid in token_ids)
+    assert token_ids[0] == 1
+
+
+@pytest.mark.asyncio
+async def test_completion_render_multiple_prompts(client):
+    """Test completion render with multiple prompts."""
+    response = await client.post(
+        "/v1/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "prompt": ["Hello world", "Goodbye world"],
+        },
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+
+    # Should return two GenerateRequest items
+    assert isinstance(data, list)
+    assert len(data) == 2
+
+    # Verify both prompts have GenerateRequest fields
+    for prompt in data:
+        assert "token_ids" in prompt
+        assert "sampling_params" in prompt
+        assert "model" in prompt
+        assert "request_id" in prompt
+        assert len(prompt["token_ids"]) > 0
+        assert prompt["request_id"].startswith("cmpl-")
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_multi_turn(client):
+    """Test chat completion render with multi-turn conversation."""
+    response = await client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "messages": [
+                {"role": "user", "content": "Hello"},
+                {"role": "assistant", "content": "Hi there!"},
+                {"role": "user", "content": "How are you?"},
+            ],
+        },
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+
+    # Verify tokenization occurred
+    assert isinstance(data, dict)
+    assert "token_ids" in data
+    assert isinstance(data["token_ids"], list)
+    assert len(data["token_ids"]) > 0
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_with_stream_true(client):
+    """Render accepts stream params but still returns JSON (non-streamed)."""
+
+    response = await client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "stream": True,
+            "stream_options": {
+                "include_usage": True,
+                "continuous_usage_stats": True,
+            },
+            "messages": [
+                {
+                    "role": "user",
+                    "content": "Stream options should be accepted by /render.",
+                }
+            ],
+        },
+    )
+
+    assert response.status_code == 200
+    assert response.headers.get("content-type", "").startswith("application/json")
+
+    data = response.json()
+    assert isinstance(data, dict)
+    assert "token_ids" in data
+    assert isinstance(data["token_ids"], list)
+    assert len(data["token_ids"]) > 0
+
+    # /render should preserve stream fields on the returned token-in request.
+    assert data.get("stream") is True
+    assert isinstance(data.get("stream_options"), dict)
+    assert data["stream_options"].get("include_usage") is True
+    assert data["stream_options"].get("continuous_usage_stats") is True
+
+
+@pytest.mark.asyncio
+async def test_completion_render_error_invalid_model(client):
+    """Test completion render with invalid model returns error."""
+    response = await client.post(
+        "/v1/completions/render",
+        json={
+            "model": "invalid-model-name",
+            "prompt": "Hello",
+        },
+    )
+
+    assert response.status_code == 404
+    data = response.json()
+    assert "error" in data
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_error_invalid_model(client):
+    """Test chat completion render with invalid model returns error."""
+    response = await client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": "invalid-model-name",
+            "messages": [{"role": "user", "content": "Hello"}],
+        },
+    )
+
+    assert response.status_code == 404
+    data = response.json()
+    assert "error" in data
+
+
+@pytest.mark.asyncio
+async def test_completion_render_no_generation(client):
+    """Verify render endpoint does not generate text."""
+    # This test verifies that calling render is fast (no generation)
+    import time
+
+    start = time.perf_counter()
+    response = await client.post(
+        "/v1/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "prompt": "Tell me a very long story about " * 10,
+        },
+    )
+    elapsed = time.perf_counter() - start
+
+    assert response.status_code == 200
+    # Render should be fast (< 1 second) since no generation
+    assert elapsed < 1.0
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_with_sampling_params(client):
+    """Verify sampling params are correctly returned by /render."""
+    response = await client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": MODEL_NAME,
+            "messages": [{"role": "user", "content": "Test sampling params"}],
+            "temperature": 0.123,
+            "top_p": 0.456,
+            "frequency_penalty": 1.1,
+        },
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+
+    assert "sampling_params" in data
+    sampling_params = data["sampling_params"]
+
+    assert sampling_params.get("temperature") == 0.123
+    assert sampling_params.get("top_p") == 0.456
+    assert sampling_params.get("frequency_penalty") == 1.1
+
+    # Check that internal fields are not present
+    assert "_all_stop_token_ids" not in sampling_params

third_party/vllm/tests/entrypoints/openai/cpu/test_render_multimodal.py

@@ -0,0 +1,155 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Multimodal tests for the /render endpoints that expose prompt preprocessing."""
+
+import httpx
+import pytest
+import pytest_asyncio
+
+from tests.utils import RemoteOpenAIServer
+from vllm.multimodal.utils import encode_image_url
+
+VISION_MODEL_NAME = "Qwen/Qwen3-VL-2B-Instruct"
+
+
+@pytest.fixture(scope="module")
+def vision_server():
+    """Vision-capable server used for multimodal /render tests."""
+
+    args = [
+        "--enforce-eager",
+        "--max-model-len",
+        "100",
+        "--max-num-seqs",
+        "1",
+        "--limit-mm-per-prompt.image",
+        "1",
+        "--limit-mm-per-prompt.video",
+        "0",
+    ]
+
+    env_overrides: dict[str, str] = {}
+
+    with RemoteOpenAIServer(
+        VISION_MODEL_NAME,
+        args,
+        env_dict=env_overrides,
+    ) as remote_server:
+        yield remote_server
+
+
+@pytest_asyncio.fixture
+async def vision_client(vision_server):
+    async with httpx.AsyncClient(
+        base_url=vision_server.url_for(""), timeout=60.0
+    ) as http_client:
+        yield http_client
+
+
+@pytest.mark.asyncio
+async def test_chat_completion_render_with_base64_image_url(
+    vision_client,
+    local_asset_server,
+):
+    """Render a multimodal chat request and verify tokens are returned."""
+
+    image = local_asset_server.get_image_asset("RGBA_comp.png")
+    data_url = encode_image_url(image, format="PNG")
+
+    assert data_url.startswith("data:image/")
+    assert ";base64," in data_url
+
+    response = await vision_client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": VISION_MODEL_NAME,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": data_url}},
+                        {"type": "text", "text": "What's in this image?"},
+                    ],
+                }
+            ],
+        },
+    )
+
+    assert response.status_code == 200
+
+    data = response.json()
+    assert isinstance(data, dict)
+    assert "token_ids" in data
+    assert isinstance(data["token_ids"], list)
+    assert len(data["token_ids"]) > 0
+
+    # Verify multimodal features are populated
+    assert "features" in data
+    features = data["features"]
+    assert features is not None
+
+    # mm_hashes: should have an "image" key with a list of hash strings
+    assert "mm_hashes" in features
+    assert "image" in features["mm_hashes"]
+    image_hashes = features["mm_hashes"]["image"]
+    assert isinstance(image_hashes, list)
+    assert len(image_hashes) > 0
+    assert all(isinstance(h, str) for h in image_hashes)
+
+    # mm_placeholders: should have an "image" key with offset/length dicts
+    assert "mm_placeholders" in features
+    assert "image" in features["mm_placeholders"]
+    image_placeholders = features["mm_placeholders"]["image"]
+    assert isinstance(image_placeholders, list)
+    assert len(image_placeholders) > 0
+    for p in image_placeholders:
+        assert "offset" in p
+        assert "length" in p
+        assert isinstance(p["offset"], int)
+        assert isinstance(p["length"], int)
+        assert p["length"] > 0
+
+
+@pytest.mark.asyncio
+async def test_tokenize_matches_render_for_multimodal_input(
+    vision_client,
+    local_asset_server,
+):
+    """`/tokenize` should match `/v1/chat/completions/render` token output."""
+
+    image = local_asset_server.get_image_asset("RGBA_comp.png")
+    data_url = encode_image_url(image, format="PNG")
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": data_url}},
+                {"type": "text", "text": "What's in this image?"},
+            ],
+        }
+    ]
+
+    render_response = await vision_client.post(
+        "/v1/chat/completions/render",
+        json={
+            "model": VISION_MODEL_NAME,
+            "messages": messages,
+        },
+    )
+    assert render_response.status_code == 200
+    render_data = render_response.json()
+
+    tokenize_response = await vision_client.post(
+        "/tokenize",
+        json={
+            "model": VISION_MODEL_NAME,
+            "messages": messages,
+        },
+    )
+    assert tokenize_response.status_code == 200
+    tokenize_data = tokenize_response.json()
+
+    assert tokenize_data["tokens"] == render_data["token_ids"]
+    assert tokenize_data["count"] == len(render_data["token_ids"])