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/test_terratorch.py

@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+
+from tests.conftest import VllmRunner
+from tests.utils import create_new_process_for_each_test
+
+
+@create_new_process_for_each_test()  # Memory is not cleaned up properly otherwise
+@pytest.mark.parametrize(
+    "model",
+    [
+        "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
+        "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-BurnScars",
+    ],
+)
+def test_inference(
+    vllm_runner: type[VllmRunner],
+    model: str,
+) -> None:
+    pixel_values = torch.full((6, 512, 512), 1.0, dtype=torch.float16)
+    location_coords = torch.full((1, 2), 1.0, dtype=torch.float16)
+    prompt = dict(
+        prompt_token_ids=[1],
+        multi_modal_data={
+            "image": {
+                "pixel_values": pixel_values,
+                "location_coords": location_coords,
+            }
+        },
+    )
+
+    with vllm_runner(
+        model,
+        runner="pooling",
+        dtype="half",
+        enforce_eager=True,
+        skip_tokenizer_init=True,
+        enable_mm_embeds=True,
+        # Limit the maximum number of sequences to avoid the
+        # test going OOM during the warmup run
+        max_num_seqs=32,
+        default_torch_num_threads=1,
+    ) as vllm_model:
+        vllm_output = vllm_model.llm.encode(prompt, pooling_task="plugin")
+        assert torch.equal(
+            torch.isnan(vllm_output[0].outputs.data).any(), torch.tensor(False)
+        )