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/samplers/test_ignore_eos.py

@@ -0,0 +1,35 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Make sure ignore_eos works.
+
+Run `pytest tests/samplers/test_ignore_eos.py`.
+"""
+
+import pytest
+
+from vllm import SamplingParams
+
+# We also test with llama because it has generation_config to specify EOS
+# (past regression).
+MODELS = ["distilbert/distilgpt2", "meta-llama/Llama-3.2-1B"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [512])
+def test_ignore_eos(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        sampling_params = SamplingParams(max_tokens=max_tokens, ignore_eos=True)
+
+        for prompt in example_prompts:
+            ignore_eos_output = vllm_model.llm.generate(
+                prompt, sampling_params=sampling_params
+            )
+            output_length = len(ignore_eos_output[0].outputs[0].token_ids)
+            assert output_length == max_tokens