fix: fix illeagel memory access (#6437)

dongjiyingdjy · web-flow · commit 17e0d0fb1add · 2025-07-31T10:01:34.000+08:00
Signed-off-by: Jiying Dong &lt;87510204+dongjiyingdjy@users.noreply.github.com&gt;
diff --git 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

Original file line number	Diff line number	Diff line change
`@@ -634,7 +634,7 @@ def __post_init__(self) -> None:`
`634`	`634`	`self.block_ids_per_seq = None`
`635`	`635`	`self.kv_block_ids_per_seq = None`
`636`	`636`	`if self.enable_flash_mla:`
`637`		`- self.block_ids_per_seq = torch.empty(`
	`637`	`+ self.block_ids_per_seq = torch.zeros(`
`638`	`638`	`[`
`639`	`639`	`self.kv_cache_manager.max_batch_size,`
`640`	`640`	`self.kv_cache_manager.max_blocks_per_seq`