]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
CANN: fix CPU memory leak in CANN backend (llama/16549)
authorChenguang Li <redacted>
Mon, 13 Oct 2025 09:01:24 +0000 (17:01 +0800)
committerGeorgi Gerganov <redacted>
Tue, 14 Oct 2025 19:07:44 +0000 (22:07 +0300)
commit01314737b5f2e45ce39e70519e1144e256f40ca9
treeb3784e00a52a70206199ae6be576a79c740e5699
parent94c04204d366b850b89a8a8302383838343ca467
CANN: fix CPU memory leak in CANN backend (llama/16549)

This commit fixes a CPU-side memory leak issue in the CANN backend,
which occurred when intermediate aclTensorList objects were not properly
released after operator execution. The leak happened during repeated
invocations of CANN ops (e.g., FlashAttention), leading to increasing
host memory usage over time.

Proper resource cleanup (aclDestroyTensorList and related release logic)
has been added to ensure that all temporary tensors are correctly freed.
src/ggml-cann/aclnn_ops.cpp