weiqingy commented on issue #889: URL: https://github.com/apache/flink-agents/issues/889#issuecomment-4966429791
Agreed, a real fix beats a retry, so let's hold #895 until we know whether the segfault can be resolved at the source. I've put the details on #893. The hardware lead looks right, but `OLLAMA_LLM_LIBRARY` won't get us there on 0.31.x, since it only filters GPU library directories now. The CPU kernel is chosen by `llama-server`, which loads the sapphirerapids ggml variant, the only one built with AMX, on those hosts. So the fix would be to stop that variant loading, in `tools/start_ollama_server.sh`. Next step I'd suggest is a small diagnostic in the integration jobs to log the CPU flags and the loaded variant, so we can confirm it rather than guess. If it holds, I'll put up the fix and close #895. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
