Closed wconstab closed 3 weeks ago
CI should pass after https://github.com/pytorch/pytorch/pull/127607 is landed.
Made a small change: In tracer mode, we don't require users to provide manual split points, because that is being taken care of by the pipeline_llama_tracer function today:
I think this is not the right way to do this. If we want the layer-split to be 'automatic', i think it should be automatic for both frontends, and we can delete the _split_points cmdline arg.
or if we want to have the cmdline arg, we should keep it behaving the same for both frontends.
I'd propose to first keep the arg for both frontends, and then do a PR that makes the cmdline arg optional and uses automation for both PRs.
Sounds good to me
Stack from ghstack (oldest at bottom):
forgot to enable tracer for tracer test in the last PR