Skip to content

Conversation

NeoZhangJianyu
Copy link
Collaborator

Revert "sycl: add usage of enqueue_functions extension (https://github.com/ggml-org/llama.cpp/pull/14244[)"](https://github.com/ggml-org/llama.cpp/commit/159e9c9df729f40029e3aec3101d64611b4d9ab9)

There are users find there is obviously memory leak with this feature.

In a inference, the memory increase 80M in one interact session (22s).
In workload test, 16GB GPU will be filled after about 2 hours.

It happens on oneAPI 2025.0- 2025.2.

Revert this PR since it includes big issue.

The issue is reported and the issue will be fixed in next official release.
After this feature is fixe in new SYCL version, this PR could be restored.

Here is the log in my test.

15:18:09.972, 0, N/A, 6008.30
15:18:10.972, 0, N/A, 6012.74
15:18:11.972, 0, N/A, 6017.19
15:18:12.972, 0, N/A, 6021.64
15:18:13.972, 0, N/A, 6026.08
15:18:14.972, 0, N/A, 6030.53
15:18:15.972, 0, N/A, 6034.97
15:18:16.972, 0, N/A, 6039.36
15:18:17.972, 0, N/A, 6043.74
15:18:18.972, 0, N/A, 6047.94
15:18:19.972, 0, N/A, 6052.20
15:18:20.972, 0, N/A, 6056.45
15:18:21.972, 0, N/A, 6060.65
15:18:22.972, 0, N/A, 6064.84
15:18:23.972, 0, N/A, 6065.28
15:18:24.972, 0, N/A, 6065.28
15:18:25.972, 0, N/A, 6065.28
15:18:26.972, 0, N/A, 6066.41
15:18:27.972, 0, N/A, 6070.67
15:18:28.972, 0, N/A, 6074.80
15:18:29.972, 0, N/A, 6079.06
15:18:30.972, 0, N/A, 6083.44
15:18:31.972, 0, N/A, 6087.39

@github-actions github-actions bot added ggml changes relating to the ggml tensor library for machine learning SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language labels Sep 10, 2025
@NeoZhangJianyu NeoZhangJianyu merged commit 704d90c into ggml-org:master Sep 12, 2025
48 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
ggml changes relating to the ggml tensor library for machine learning SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants