Are there any good papers exploring precision schedules for training? Something like higher precision initially, than fp8 or lower during the meat of training and then maybe back up for the end? or maybe even lower precision for the end?

Training at 2K scale on Crusoe B200 GPUs shows 1.22x–1.28x acceleration using TorchAO’s MXFP8 implementation on TorchTitan, with equivalent convergence to BF16. 🔗 Read our latest blog to learn more: hubs.la/Q03GHFnc0 #PyTorch #OpenSourceAI #AIInfrastructure

Having claude code iterate on slash commands w/ me feels so wrong

:)

Google utilizing tanh softcapping for its models perfectly encapsulates the differences between gpus and tpus

I caved and purchased chatgpt plus for the month.. ehh I think I am still team orange. Fits my style of plan mode, iterate iterate iterate -> implement

blog.demofox.org/2017/11/21/flo… A very good table

github.com/pytorch/pytorc… I tried my best to explain this here

I actually still really like deepwiki and am dumbfounded that GitHub hasn’t bear hugged this product space

github.com/pytorch/pytorc… Major W for claude. Could I have bisected this env difference? Yes. Did I give it a good starting point to work from? Yes. But still going dangerously skip and sitting back was alot easier on a Friday

"I can't directly modify the formatting of the cells in your document." WTF IS YOUR PURPOSE THEN...

Someone make my day :) github.com/meta-pytorch/a…

Claude is my worker Codex is my explorer GPT5 is my scholar

github.com/microsoft/vsco… please give me more tab completion options. I dont like cursor but I want their tab model

RE: Stacked Diffs on @GitHub After discussion w Taylor Blau, we can implement stacked PRs/PR groups already (in fact we kind of do with Copilot) but restacking (automatically fanning out changes from the bottom of the the stack upwards) would be wildly inefficient. To do it

Actually, if it’s not based off of commits I’m a little bearish

UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'.

The San Francisco billboards are the strongest evidence I have seen for AI being a bubble