Closed
【缺陷报告】内存泄漏 - 文件test_aclnn_moe_token_permute_with_routing_map_grad.cpp - 函数main - 行号238 #4810
zhangjunkai9Add on  Aug 31Closed at  Sep 12
zhangjunkai9
Aug 31 Create

缺陷信息

缺陷类型memoryLeak (内存泄漏)
函数main
文件moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp
行号238
置信度95%
分析模型maas-glm-5.1-zhipu

缺陷描述

函数 main 在第158-203行通过 CreateAclTensor 分配了6个 aclTensor 及对应的 device 内存(permuted_output_grad/permuted_output_grad_Addr、ppermutedProbsOutputGrad/permutedProbsOutputGrad_Addr、sortedIndices/sortedIndicesAddr、proutingMapOptional/routingMapOptionalAddr、out/outAddr、out2/outAddr2)。但在第237-245行的释放清理代码中,仅释放了 permuted_output_grad、sortedIndices、out 三个 tensor 及其 device 地址,遗漏了 ppermutedProbsOutputGrad、proutingMapOptional、out2 三个 tensor 及对应的 permutedProbsOutputGrad_Addr、routingMapOptionalAddr、outAddr2 三块 device 内存,造成内存泄漏。

事实核查

经复核确认(1):第158-203行通过CreateAclTensor分配了6个aclTensor及对应device内存(permuted_output_grad/ppermutedProbsOutputGrad/sortedIndices/proutingMapOptional/out/out2),但第237-245行释放代码仅aclDestroyTensor了permuted_output_grad、sortedIndices、out三个tensor,aclrtFree了permuted_output_grad_Addr、sortedIndicesAddr、outAddr三块地址,遗漏了ppermutedProbsOutputGrad、proutingMapOptional、out2三个tensor及permutedProbsOutputGrad_Addr、routingMapOptionalAddr、outAddr2三块device内存,模式成立。

(2):无任何防护,cleanup代码直接遗漏。

(3):main为入口函数,可达。

(4):该泄漏在正常路径上必然发生——无需任何错误条件即到达不完整的cleanup代码,触发可满足。

确认缺陷。

数据流证据

Source(问题源头)

moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp:166-167 行 CreateAclTensor(..., &permutedProbsOutputGrad_Addr, ..., &ppermutedProbsOutputGrad) 分配 ppermutedProbsOutputGrad tensor 和 permutedProbsOutputGrad_Addr device 内存(分配点之一)

Sink(问题爆发点)

moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp:237-245 行 释放清理代码遗漏 ppermutedProbsOutputGrad/proutingMapOptional/out2 及其 device 地址(泄漏 sink)

传播路径:

# 文件 行号 说明
1 moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp 166-167 CreateAclTensor 分配 ppermutedProbsOutputGrad 和 permutedProbsOutputGrad_Addr(source)
2 moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp 184-185 CreateAclTensor 分配 proutingMapOptional 和 routingMapOptionalAddr(source)
3 moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp 202-203 CreateAclTensor 分配 out2 和 outAddr2(source)
4 moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp 238-240 aclDestroyTensor 仅销毁 permuted_output_grad、sortedIndices、out,遗漏 ppermutedProbsOutputGrad、proutingMapOptional、out2
5 moe/moe_token_permute_with_routing_map_grad/examples/test_aclnn_moe_token_permute_with_routing_map_grad.cpp 243-245 aclrtFree 仅释放 permuted_output_grad_Addr、sortedIndicesAddr、outAddr,遗漏 permutedProbsOutputGrad_Addr、routingMapOptionalAddr、outAddr2(sink)

调用链

可达调用链1 起点:(entry functions) → 终点:main 深度:0

# 文件 函数 函数起始行 调用点行
1 attention/attention_update/examples/test_aclnn_attention_update.cpp main 77 —

修复建议

    // 6. 释放aclTensor和aclScalar
    aclDestroyTensor(permuted_output_grad);
    aclDestroyTensor(ppermutedProbsOutputGrad);
    aclDestroyTensor(sortedIndices);
    aclDestroyTensor(proutingMapOptional);
    aclDestroyTensor(out);
    aclDestroyTensor(out2);

    // 7. 释放device资源
    aclrtFree(permuted_output_grad_Addr);
    aclrtFree(permutedProbsOutputGrad_Addr);
    aclrtFree(sortedIndicesAddr);
    aclrtFree(routingMapOptionalAddr);
    aclrtFree(outAddr);
    aclrtFree(outAddr2);
likedislike
weihao18Collaborators
Sep 1 Comments:

你好,建议汇总到具体算子进行集中处理

likedislike
游震Collaborators
Sep 3 Comments:

/assign @guoqiuhao

likedislike
CANN-robotCANN-robotCollaborators
Sep 3 Assign guoqiuhao as responsible person
游游震Collaborators
Sep 3 Assign cc-z as responsible person
游游震Collaborators
Sep 3 Removed responsible person guoqiuhao
Kkdy18482276080Collaborators
Sep 3 Assign kdy18482276080 as responsible person
Kkdy18482276080Collaborators
Sep 3 Removed responsible person kdy18482276080
Kkdy18482276080Collaborators
Sep 3 Assign kdy18482276080 as responsible person
Kkdy18482276080Collaborators
Sep 3 Associated pull request
Kkdy18482276080Collaborators
Sep 3 Removed responsible person cc-z
CANN-robotCANN-robotCollaborators
Sep 12 Close issue
CANN-robotCANN-robotCollaborators
Sep 12 Added label: resolved