wozeparrot
wozeparrot
Tracking at #1698
Closing as the code that this touches has been long changed by refactors.
Just wanted an update on this. How's it coming along?
1. We can just do a subset of the tests, just the really important ones. 2. Using the same codegen should be doable, and the `kernel_launcher` thing shouldn't be needed,...
This does make it so that core tinygrad is importing from `extra/` do we want to move the bindings over or is it fine if we keep it in `extra/`...
That makes sense, I may depend on this just to get p2p transfers to work as HIP does support them.
Currently doing another run cuz I found that the metric code was wrong
ok, I got a word error rate of 7.38024337340539% across the filtered dev-clean set that mlperf specifies. the mlperf word error rate is specified as 7.452253714852645%.
> Oh, test is failing in LLVM + GPU. GPU test failure doesn't look related, will take a look at llvm
Both tests should be fixed now