[ModelRunner V2] Don't pin reused flashinfer tensors (#32799)

Signed-off-by: Nick Hill <nickhill123@gmail.com>

vllm/v1/attention/backends/flashinfer.py

@@ -603,7 +603,12 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 "earlier GPUs."
             )
         # Preparing persistent buffers
-        self.pin_memory = is_pin_memory_available()
+        # Since we do not have explicit synchronization in ModelRunnerV2, we do not pin
+        # reused CPU buffers to avoid a race condition between step N async copies to
+        # GPU and step N+1 buffer updates.
+        self.pin_memory = (
+            not envs.VLLM_USE_V2_MODEL_RUNNER and is_pin_memory_available()
+        )
         self.paged_kv_indptr = self._make_buffer(max_num_reqs + 1)
         self.paged_kv_indptr_cpu_buffer = torch.zeros_like(
             self.paged_kv_indptr.cpu, pin_memory=self.pin_memory