diff --git a/tensorrt_llm/_torch/attention/backends/trtllm.py b/tensorrt_llm/_torch/attention/backends/trtllm.py
index 21ca568ab81d..857737edc4f0 100644
--- a/tensorrt_llm/_torch/attention/backends/trtllm.py
+++ b/tensorrt_llm/_torch/attention/backends/trtllm.py
@@ -944,6 +944,11 @@ def prepare_flash_mla(self) -> None:
         block_ids_per_seq = maybe_pin_memory(
             self.kv_cache_manager.get_block_ids_per_seq(self.request_ids))
         num_blocks = block_ids_per_seq.shape[1]
+        # Retain the source CPU buffer so it outlives the non-blocking H2D
+        # copies below; back-to-back prepare_flash_mla calls during CUDA-graph
+        # capture can otherwise free it before the DMA completes, leaving
+        # garbage entries that trigger an OOB read in the FlashMLA kernel.
+        self._flash_mla_src_block_ids_cpu = block_ids_per_seq
         self.kv_block_ids_per_seq.fill_(0)
         self.kv_block_ids_per_seq[:self.num_seqs, :num_blocks].copy_(
             block_ids_per_seq, non_blocking=True)