merrymercy edited a comment on pull request #6671: URL: https://github.com/apache/incubator-tvm/pull/6671#issuecomment-708845921
Autotvm uses a lot of python multiprocessing and I expect it will be much slower when using spawn. AutoTVM uses multiprocessing for feature extraction. So it needs to launch about 50,000 tasks every measurement batch. Ansor does not rely on multiprocessing too much. So the situation for Ansor is better. However, I'd like to see benchmark numbers about fork vs. spawn. i.e. Run a search with `n-trials=64` with (Ansor, AutoTVM) x (fork, spawn) X (CPU, GPU) (Optional). The requirement from my side is that I don't want to see any performance regression. I prefer to only use the slower and safer fallback mechanism when the fast approach does not work. Removing all multiprocessing requires huge engineering effort. ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
