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/csrc/cuda_utils_kernels.cu

@@ -0,0 +1,35 @@
+#include "cuda_utils.h"
+#ifdef USE_ROCM
+  #include <hip/hip_runtime.h>
+  #include <hip/hip_runtime_api.h>
+#endif
+
+int64_t get_device_attribute(int64_t attribute, int64_t device_id) {
+  // Return the cached value on subsequent calls
+  static int value = [=]() {
+    int device = static_cast<int>(device_id);
+    if (device < 0) {
+      CUDA_CHECK(cudaGetDevice(&device));
+    }
+    int value;
+    CUDA_CHECK(cudaDeviceGetAttribute(
+        &value, static_cast<cudaDeviceAttr>(attribute), device));
+    return static_cast<int>(value);
+  }();
+
+  return value;
+}
+
+int64_t get_max_shared_memory_per_block_device_attribute(int64_t device_id) {
+  int64_t attribute;
+  // https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
+  // cudaDevAttrMaxSharedMemoryPerBlockOptin = 97 if not is_hip() else 74
+
+#ifdef USE_ROCM
+  attribute = hipDeviceAttributeMaxSharedMemoryPerBlock;
+#else
+  attribute = cudaDevAttrMaxSharedMemoryPerBlockOptin;
+#endif
+
+  return get_device_attribute(attribute, device_id);
+}