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/quantization/gptq/qdq_8.cuh

@@ -0,0 +1,30 @@
+/*
+Copied from https://github.com/turboderp/exllamav2
+*/
+
+#ifndef _qdq_8_cuh
+#define _qdq_8_cuh
+
+#include "qdq_util.cuh"
+
+namespace vllm {
+namespace gptq {
+
+__forceinline__ __device__ void shuffle_8bit_4(uint32_t* q, int stride) {}
+
+__forceinline__ __device__ void dequant_8bit_8(const uint32_t q_0,
+                                               const uint32_t q_1,
+                                               half2 (&dq)[4], int stride,
+                                               const uint32_t zero) {
+  half dqh[8];
+  for (int i = 0; i < 4; i++) dqh[i] = dq_ns(exb(q_0, i * 8, 0xff), zero);
+  for (int i = 0; i < 4; i++) dqh[i + 4] = dq_ns(exb(q_1, i * 8, 0xff), zero);
+
+  for (int i = 0; i < 4; i++)
+    dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
+}
+
+}  // namespace gptq
+}  // namespace vllm
+
+#endif