Profiling oddity - why so slow sometimes?

[evanmiller]

Often when I attempt to profile the llama.cpp main process using Instruments, the process drags to a crawl with a vast majority of the time spent in ggml_graph_compute_thread:

I cannot reliably reproduce the condition. I was wondering if anyone had seen similar issues and whether there might be a workaround.

[evanmiller]

Seems to have been resolved by recent commits; I suspect it had to do with the multi-threading in BLAS interfering with the GGML scheduling model.