asu
asu
The landscape of things have changed since. If I do any more progress on y8 it probably won't be on the RP2040 but on the RP2350 which solves a number...
We could benefit from using the Lua test suite over the y8 fork, too.
Squashed the commit into one so that the removed pretrained HF model stuff doesn't make it to the history; looking into the other CI issues (which seems to be because...
We will both be somewhat unavailable/busy in the near future but I will try to review by the end of this month.
~~pre-commit fail resolution depends on #2665 merge~~ done
I'll get to continuing changes when I have news about getting EPAC/REPERE at the lab in their original format for reproducing this.
@pchampio could you send me the file list/hierarchy of those datasets? I have had discussions at the lab and there is some confusion about how e.g. REPERE has been obtained...
No problem, ty :)
Can't repro on Jean Zay with 4x A100 on DDP either with PyTorch 2.3.0
Can't repro on Adastra with 8x MI250X on DDP either, also PyTorch 2.3.0 Also: > From the doc of 2.3, we can read: "ProcessGroupNCCL now relies on stream synchronization instead...