unavailableun
unavailableun
> we are waiting on [pytorch/pytorch#124885](https://github.com/pytorch/pytorch/pull/124885) to merge. cc: @pruthvistony @xinyazhang @amathews-amd @fxmarty Looks like this one has been merged. Any plan to re-enable SDPA on ROCM?
Sure, need some time to prepare a clean repro script.
@therealnaveenkamal Here is the repro script: https://gist.github.com/unavailableun/5efab330f82ec6af051717adee4c3455 Running command on my local device (1 node with 2 A6000 GPUs): `torchrun --nproc_per_node=2 train.py` Environment: - cuda 12.4 - python 3.10 -...
Above MemoryProfiling error was throwed from below memory profiling part in `compile/backend.py`, seems no big impact to running, so I commented this part, while still get an aot TypeError. ...
Hi @tohtana Have you reproduced the error after your fix?
I saw alert only for CP & FlexAttention, does that mean CP works with SDPA?