Maxime

For a couple reasons: - Building is slower with vsxmake - Cleaning my solution takes almost 2 minutes Everytime I give vsxmake a try again because I encounter an issue...

Here is a train run on Gemma 2B, sadly laptop went to sleep right before the end but this is a 14 hours run, should be representative enough. [log_1715964526.txt](https://github.com/pytorch/torchtune/files/15361933/log_1715964526.txt)

We need to find people who had the issue and have them test the feature changes before merging.

No, we do care but it's rather low priority