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/cuda/scripts/check_device_count_respects_env.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Check that device_count respects CUDA_VISIBLE_DEVICES after platform import."""
+
+import os
+import sys
+
+for key in ["CUDA_VISIBLE_DEVICES", "HIP_VISIBLE_DEVICES", "ROCR_VISIBLE_DEVICES"]:
+    os.environ.pop(key, None)
+
+import torch  # noqa: E402
+
+from vllm.platforms import current_platform  # noqa: F401, E402
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+count = torch.accelerator.device_count()
+
+if count == 0:
+    sys.exit(0)  # Skip: no GPUs available
+
+assert count == 1, f"device_count()={count}, expected 1"
+print("OK")

third_party/vllm/tests/cuda/scripts/check_platform_no_cuda_init.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Check that vllm.platforms import does not initialize CUDA."""
+
+import os
+
+for key in ["CUDA_VISIBLE_DEVICES", "HIP_VISIBLE_DEVICES", "ROCR_VISIBLE_DEVICES"]:
+    os.environ.pop(key, None)
+
+import torch  # noqa: E402
+
+assert not torch.cuda.is_initialized(), "CUDA initialized before import"
+
+from vllm.platforms import current_platform  # noqa: E402
+
+assert not torch.cuda.is_initialized(), (
+    f"CUDA was initialized during vllm.platforms import on {current_platform}"
+)
+print("OK")

third_party/vllm/tests/cuda/test_cuda_compatibility_path.py

@@ -0,0 +1,187 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for CUDA forward compatibility path logic in env_override.py.
+
+Verifies the opt-in LD_LIBRARY_PATH manipulation for CUDA compat libs,
+including env var parsing, path detection, and deduplication.
+"""
+
+import os
+from unittest.mock import patch
+
+import pytest
+
+# Import the functions directly (they're module-level in env_override)
+# We must import them without triggering the module-level side effects,
+# so we import the functions by name after the module is already loaded.
+from vllm.env_override import (
+    _get_torch_cuda_version,
+    _maybe_set_cuda_compatibility_path,
+)
+
+
+class TestCudaCompatibilityEnvParsing:
+    """Test VLLM_ENABLE_CUDA_COMPATIBILITY env var parsing."""
+
+    def test_disabled_by_default(self, monkeypatch):
+        """Compat path is NOT set when env var is absent."""
+        monkeypatch.delenv("VLLM_ENABLE_CUDA_COMPATIBILITY", raising=False)
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        assert (
+            "LD_LIBRARY_PATH" not in os.environ
+            or os.environ.get("LD_LIBRARY_PATH", "") == ""
+        )
+
+    @pytest.mark.parametrize("value", ["0", "false", "False", "no", ""])
+    def test_disabled_values(self, monkeypatch, value):
+        """Various falsy values should not activate compat path."""
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", value)
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        # LD_LIBRARY_PATH should not be set (or remain empty)
+        ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+        assert "compat" not in ld_path
+
+    @pytest.mark.parametrize("value", ["1", "true", "True", " 1 ", " TRUE "])
+    def test_enabled_values_with_valid_path(self, monkeypatch, tmp_path, value):
+        """Truthy values activate compat path when a valid path exists."""
+        compat_dir = tmp_path / "compat"
+        compat_dir.mkdir()
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", value)
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(compat_dir))
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+        assert str(compat_dir) in ld_path
+
+
+class TestCudaCompatibilityPathDetection:
+    """Test path detection: custom override, conda, default."""
+
+    def test_custom_path_override(self, monkeypatch, tmp_path):
+        """VLLM_CUDA_COMPATIBILITY_PATH takes highest priority."""
+        custom_dir = tmp_path / "my-compat"
+        custom_dir.mkdir()
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(custom_dir))
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+        assert ld_path.startswith(str(custom_dir))
+
+    def test_conda_prefix_fallback(self, monkeypatch, tmp_path):
+        """Falls back to $CONDA_PREFIX/cuda-compat if custom not set."""
+        conda_dir = tmp_path / "conda-env"
+        compat_dir = conda_dir / "cuda-compat"
+        compat_dir.mkdir(parents=True)
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.delenv("VLLM_CUDA_COMPATIBILITY_PATH", raising=False)
+        monkeypatch.setenv("CONDA_PREFIX", str(conda_dir))
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+        assert str(compat_dir) in ld_path
+
+    def test_no_valid_path_does_nothing(self, monkeypatch):
+        """When enabled but no valid path exists, LD_LIBRARY_PATH unchanged."""
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", "/nonexistent/path")
+        monkeypatch.delenv("CONDA_PREFIX", raising=False)
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        with patch("vllm.env_override._get_torch_cuda_version", return_value=None):
+            _maybe_set_cuda_compatibility_path()
+        assert os.environ.get("LD_LIBRARY_PATH", "") == ""
+
+    def test_default_cuda_path_fallback(self, monkeypatch, tmp_path):
+        """Falls back to /usr/local/cuda-{ver}/compat via torch version."""
+        fake_cuda = tmp_path / "cuda-12.8" / "compat"
+        fake_cuda.mkdir(parents=True)
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.delenv("VLLM_CUDA_COMPATIBILITY_PATH", raising=False)
+        monkeypatch.delenv("CONDA_PREFIX", raising=False)
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        with (
+            patch("vllm.env_override._get_torch_cuda_version", return_value="12.8"),
+            patch(
+                "vllm.env_override.os.path.isdir",
+                side_effect=lambda p: p == "/usr/local/cuda-12.8/compat"
+                or os.path.isdir(p),
+            ),
+        ):
+            _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+        assert "/usr/local/cuda-12.8/compat" in ld_path
+
+
+class TestCudaCompatibilityLdPathManipulation:
+    """Test LD_LIBRARY_PATH prepend and deduplication logic."""
+
+    def test_prepends_to_empty_ld_path(self, monkeypatch, tmp_path):
+        """Compat path is set when LD_LIBRARY_PATH is empty."""
+        compat_dir = tmp_path / "compat"
+        compat_dir.mkdir()
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(compat_dir))
+        monkeypatch.delenv("LD_LIBRARY_PATH", raising=False)
+        _maybe_set_cuda_compatibility_path()
+        assert os.environ["LD_LIBRARY_PATH"] == str(compat_dir)
+
+    def test_prepends_to_existing_ld_path(self, monkeypatch, tmp_path):
+        """Compat path is prepended before existing entries."""
+        compat_dir = tmp_path / "compat"
+        compat_dir.mkdir()
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(compat_dir))
+        monkeypatch.setenv("LD_LIBRARY_PATH", "/usr/lib:/other/lib")
+        _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ["LD_LIBRARY_PATH"]
+        parts = ld_path.split(os.pathsep)
+        assert parts[0] == str(compat_dir)
+        assert "/usr/lib" in parts
+        assert "/other/lib" in parts
+
+    def test_deduplicates_existing_compat_path(self, monkeypatch, tmp_path):
+        """If compat path already in LD_LIBRARY_PATH, move to front."""
+        compat_dir = tmp_path / "compat"
+        compat_dir.mkdir()
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(compat_dir))
+        monkeypatch.setenv(
+            "LD_LIBRARY_PATH",
+            f"/usr/lib:{compat_dir}:/other/lib",
+        )
+        _maybe_set_cuda_compatibility_path()
+        ld_path = os.environ["LD_LIBRARY_PATH"]
+        parts = ld_path.split(os.pathsep)
+        assert parts[0] == str(compat_dir)
+        assert parts.count(str(compat_dir)) == 1
+
+    def test_already_at_front_is_noop(self, monkeypatch, tmp_path):
+        """If compat path is already first, don't modify LD_LIBRARY_PATH."""
+        compat_dir = tmp_path / "compat"
+        compat_dir.mkdir()
+        original = f"{compat_dir}:/usr/lib"
+        monkeypatch.setenv("VLLM_ENABLE_CUDA_COMPATIBILITY", "1")
+        monkeypatch.setenv("VLLM_CUDA_COMPATIBILITY_PATH", str(compat_dir))
+        monkeypatch.setenv("LD_LIBRARY_PATH", original)
+        _maybe_set_cuda_compatibility_path()
+        assert os.environ["LD_LIBRARY_PATH"] == original
+
+
+class TestGetTorchCudaVersion:
+    """Test _get_torch_cuda_version() helper."""
+
+    def test_returns_string_when_torch_available(self):
+        """Should return a CUDA version string like '12.8'."""
+        version = _get_torch_cuda_version()
+        # torch is installed in vllm's environment
+        assert version is None or isinstance(version, str)
+
+    def test_returns_none_when_torch_missing(self):
+        """Should return None when torch is not importable."""
+        with patch(
+            "vllm.env_override.importlib.util.find_spec",
+            return_value=None,
+        ):
+            assert _get_torch_cuda_version() is None

third_party/vllm/tests/cuda/test_cuda_context.py

@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import ctypes
+from concurrent.futures import ThreadPoolExecutor
+
+import pytest
+import torch
+
+from vllm.platforms import current_platform
+
+
+def check_cuda_context():
+    """Check CUDA driver context status"""
+    try:
+        cuda = ctypes.CDLL("libcuda.so")
+        device = ctypes.c_int()
+        result = cuda.cuCtxGetDevice(ctypes.byref(device))
+        return (True, device.value) if result == 0 else (False, None)
+    except Exception:
+        return False, None
+
+
+def run_cuda_test_in_thread(device_input, expected_device_id):
+    """Run CUDA context test in separate thread for isolation"""
+    try:
+        # New thread should have no CUDA context initially
+        valid_before, device_before = check_cuda_context()
+        if valid_before:
+            return (
+                False,
+                "CUDA context should not exist in new thread, "
+                f"got device {device_before}",
+            )
+
+        # Test setting CUDA context
+        current_platform.set_device(device_input)
+
+        # Verify context is created correctly
+        valid_after, device_id = check_cuda_context()
+        if not valid_after:
+            return False, "CUDA context should be valid after set_cuda_context"
+        if device_id != expected_device_id:
+            return False, f"Expected device {expected_device_id}, got {device_id}"
+
+        return True, "Success"
+    except Exception as e:
+        return False, f"Exception in thread: {str(e)}"
+
+
+class TestSetCudaContext:
+    """Test suite for the set_cuda_context function."""
+
+    @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA not available")
+    @pytest.mark.parametrize(
+        argnames="device_input,expected_device_id",
+        argvalues=[
+            (0, 0),
+            (torch.device("cuda:0"), 0),
+            ("cuda:0", 0),
+        ],
+        ids=["int", "torch_device", "string"],
+    )
+    def test_set_cuda_context_parametrized(self, device_input, expected_device_id):
+        """Test setting CUDA context in isolated threads."""
+        with ThreadPoolExecutor(max_workers=1) as executor:
+            future = executor.submit(
+                run_cuda_test_in_thread, device_input, expected_device_id
+            )
+            success, message = future.result(timeout=30)
+        assert success, message
+
+    @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA not available")
+    def test_set_cuda_context_invalid_device_type(self):
+        """Test error handling for invalid device type."""
+        with pytest.raises(ValueError, match="Expected a cuda device"):
+            current_platform.set_device(torch.device("cpu"))
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])

third_party/vllm/tests/cuda/test_platform_no_cuda_init.py

@@ -0,0 +1,48 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Test that platform imports do not prematurely initialize CUDA.
+
+This is critical for Ray-based multi-GPU setups where workers need to
+set CUDA_VISIBLE_DEVICES after importing vLLM but before CUDA is initialized.
+If CUDA is initialized during import, device_count() gets locked and ignores
+subsequent env var changes.
+"""
+
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+SCRIPTS_DIR = Path(__file__).parent / "scripts"
+
+
+def run_script(script_name: str) -> subprocess.CompletedProcess:
+    """Run a test script in a subprocess with clean CUDA state."""
+    script_path = SCRIPTS_DIR / script_name
+    return subprocess.run(
+        [sys.executable, str(script_path)],
+        capture_output=True,
+        text=True,
+    )
+
+
+def test_platform_import_does_not_init_cuda():
+    """Test that importing vllm.platforms does not initialize CUDA."""
+    result = run_script("check_platform_no_cuda_init.py")
+    if result.returncode != 0:
+        pytest.fail(f"Platform import initialized CUDA:\n{result.stderr}")
+
+
+def test_device_count_respects_env_after_platform_import():
+    """Test that device_count respects CUDA_VISIBLE_DEVICES after import."""
+    result = run_script("check_device_count_respects_env.py")
+    if result.returncode != 0:
+        pytest.fail(
+            f"device_count does not respect env var after import:\n{result.stderr}"
+        )
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])