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/models/multimodal/processing/test_idefics3.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for Idefics3's multimodal preprocessing kwargs."""
+
+import pytest
+from packaging.version import Version
+from transformers import Idefics3Config
+from transformers import __version__ as TRANSFORMERS_VERSION
+
+from vllm.multimodal import MULTIMODAL_REGISTRY
+
+from ....conftest import ImageTestAssets
+from ...utils import build_model_context
+
+
+@pytest.mark.skipif(
+    Version(TRANSFORMERS_VERSION) < Version("5.2.0"),
+    reason="See https://github.com/huggingface/transformers/pull/43948",
+)
+@pytest.mark.parametrize("model_id", ["HuggingFaceM4/Idefics3-8B-Llama3"])
+@pytest.mark.parametrize(
+    ("mm_processor_kwargs", "expected_toks_per_img"),
+    [
+        ({"size": {"longest_edge": 364}}, 169),
+        ({"size": {"longest_edge": 728}}, 169 * (2**2 + 1)),
+    ],
+)
+@pytest.mark.parametrize("num_imgs", [1, 2])
+@pytest.mark.parametrize("kwargs_on_init", [True, False])
+def test_processor_override(
+    image_assets: ImageTestAssets,
+    model_id: str,
+    mm_processor_kwargs: dict[str, object],
+    expected_toks_per_img: int,
+    num_imgs: int,
+    kwargs_on_init: bool,
+):
+    """Ensure Idefics3MultiModalProcessor handles num_crops properly."""
+    # Same as the previous test - don't initialize mm_processor_kwargs
+    # in this test and assume that the kwargs will be correctly expanded by
+    # the partial when calling the custom input processor.
+    ctx = build_model_context(
+        model_id,
+        mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+    hf_processor_mm_kwargs = {} if kwargs_on_init else mm_processor_kwargs
+
+    # Build the image str / prompt based on the number of images we pass
+    placeholders = (
+        "<image>"
+        if num_imgs == 1
+        else "\n".join(f"Image-{i}: <image>\n" for i in range(1, num_imgs + 1))
+    )
+    prompt = f"<|begin_of_text|>User:{placeholders}\n<end_of_utterance>\nAssistant:"  # noqa: E501
+
+    # Build mm_data
+    image_size = ctx.get_hf_config(Idefics3Config).vision_config.image_size
+    dummy_image_size = (image_size * 4, image_size * 4)
+    dummy_image = image_assets[0].pil_image.resize(dummy_image_size)
+    mm_data = {"image": [dummy_image] * num_imgs}
+
+    processed_inputs = processor(
+        prompt,
+        mm_items=processor.info.parse_mm_data(mm_data),
+        hf_processor_mm_kwargs=hf_processor_mm_kwargs,
+    )
+
+    # Ensure the placeholders format are correct
+    hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
+    hf_processed_inputs = hf_processor(
+        text=prompt,
+        images=mm_data["image"],
+        **processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs),
+    )
+    assert processed_inputs["prompt_token_ids"] == hf_processed_inputs["input_ids"][0]
+
+    # Ensure we have the right number of placeholders per num_crops size
+    image_token_id = ctx.get_hf_config().image_token_id
+    img_tok_count = processed_inputs["prompt_token_ids"].count(image_token_id)
+    assert img_tok_count == expected_toks_per_img * num_imgs