qazal
qazal
left suggestions with the unit tests in #4220 - in general: 1. Agreed with geohot, linearizer changes should factorize with tests - I already found a weird flakiness in AST...
Yeah! This is very much appreciated. I don't think anyone is taking this so feel free to own the progress.
thank you! "external/process_replay/master_schedule.py" is in .gitignore, we only use it for process replay. 1 PR per test file sounds good.
I don't like that is is calling type_verify again, it should be all in `__post_init__` (or just merge into type_verify)
we really don't wanna do this in a toposort, can you try doing it with graph rewrite rules?
hmm so now src[3] can mean two different things... This does change the spec yea. If this is the only way to do it, the multiple temps option makes more...
lgtm
picking this back up after 6160
Looks good! let's move forward with this diff after benchmarks are green.
This is making llama somewhat slower:  and mixtral too:  I suspect it's the compiler because you change from temp1 to temp0.