[CI][Benchmarks] update llama.cpp and requirements to latest #17881
Conversation
This patch updates llama.cpp to the latest available version, uses a new, more relevant, GGUF model, and updates oneAPI to 2025.1. I was trying to avoid updating oneAPI, but the latest llama.cpp internal pooling logic seems to be broken on 2025.0, resulting in double-free errors when using older oneAPI components. The utils.download function also had to be updated, because it was using a deprecated features and didn't work on some configurations.
|
@intel/llvm-gatekeepers please merge. The CI failure is unrelated (system is dead). |
|
Just in case - if some check failed due to sporadic failure in CI, it is better to restart it. If CI failure happens more that one time, there should be a GH issue linked to this PR. |
It's not black and white. If CI is heavily loaded and people would start doing such restarts that won't help anybody. |
Right, in this case restarting would only put additional strain on CI. These scripts don't touch sycl or its tests. Only benchmark CI scripts that currently run on a separate infrastructure. |
It's not about that patch -- if something is broken in CI, we need to identify it and report it. |
This patch updates llama.cpp to the latest available version, uses a new, more relevant, GGUF model, and updates oneAPI to 2025.1.
I was trying to avoid updating oneAPI, but the latest llama.cpp internal pooling logic seems to be broken on 2025.0, resulting in double-free errors when using older oneAPI components.
The utils.download function also had to be updated, because it was using a deprecated features and didn't work on some configurations.