[BUGFIX] Fix `_dummy_run` missing `prepare_inputs_event` synchronization (#34866)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>

vllm/v1/worker/gpu_model_runner.py

@@ -4771,8 +4771,13 @@ class GPUModelRunner(
             ubatch_slices=ubatch_slices_padded,
         )
 
-        # If force_attention is True, we always capture attention. Otherwise,
-        # it only happens for cudagraph_runtime_mode=FULL.
+        # _dummy_run shares pinned CPU buffers (seq_lens, query_start_loc,
+        # etc.) with execute_model.  It must participate in the same event
+        # protocol so that back-to-back dummy/real steps don't overwrite
+        # pinned memory while a prior non_blocking H2D DMA is still reading.
+        with self.synchronize_input_prep():
+            # If force_attention is True, we always capture attention.
+            # Otherwise, it only happens for cudagraph_runtime_mode=FULL.
             if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
                 if create_mixed_batch:
                     # In the mixed batch mode (used for FI warmup), we use
@@ -4795,7 +4800,7 @@ class GPUModelRunner(
                     num_tokens_padded=num_tokens_padded if pad_attn else None,
                     num_reqs=num_reqs_padded,
                     max_query_len=max_query_len,
-                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
+                    ubatch_slices=(ubatch_slices_padded if pad_attn else ubatch_slices),
                     for_cudagraph_capture=is_graph_capturing,
                     slot_mappings=slot_mappings_by_group,
                 )