Gregory Comer
Gregory Comer
I'm seeing some path length issues when building tokenizers in-place. I'm going to see if I can set up a symlink in the windows setup CI script.
Rebasing to test with https://github.com/pytorch/executorch/pull/13923 to be safe, but will land today if everything looks good.
Removing 1.0 milestone from this PR in favor of using the pip package, when available. I'd need to make some more aggressive infra changes to the wheel build to make...
It looks like we are missing an operator implementation for this op. @manuelcandales can we add this to our op requests? Alternatively, we could request a decomp in export.
Hi @tdasika17, As a first step, can you try a few things: - Switch from to_backend() to to_edge_transform_and_lower(). Here's an example: https://pytorch.org/executorch/0.6/backends-xnnpack.html#using-the-xnnpack-backend. - Make sure you are testing with a...
@MartinPavella @StrycekSimon I'm looking at optimizing out clones more aggressively from the graph during to_edge. I see that the NXP backend has some test coverage looking at clones in the...
@fbarchard Here's the unaligned load fix.
Manual pick to update torchcodec nightly pin: https://github.com/pytorch/executorch/pull/15879
Use the release torchao version with OpenVINO: https://github.com/pytorch/executorch/pull/15880
One additional note is that aliasing analysis feels pretty fragile as is. I fixed several subtle issues (luckily caught by CI) where my changes were accidentally planning two seperate tensors...