diff --git a/tensorrt_llm/_torch/attention_backend/trtllm.py b/tensorrt_llm/_torch/attention_backend/trtllm.py
index 143fae88d62..6d4235734a9 100644
--- a/tensorrt_llm/_torch/attention_backend/trtllm.py
+++ b/tensorrt_llm/_torch/attention_backend/trtllm.py
@@ -634,7 +634,7 @@ def __post_init__(self) -> None:
             self.block_ids_per_seq = None
             self.kv_block_ids_per_seq = None
             if self.enable_flash_mla:
-                self.block_ids_per_seq = torch.empty(
+                self.block_ids_per_seq = torch.zeros(
                     [
                         self.kv_cache_manager.max_batch_size,
                         self.kv_cache_manager.max_blocks_per_seq