Closed VibhuJawa closed 3 weeks ago
Currently our memory estimation works only with torch's default memory allocator, we should make sure it works with RMM allocator plugged in too.
torch's
Like ensure below works for RMM backed Pytorch allocator too. https://github.com/rapidsai/crossfit/blob/6c213c96b3366298c5681c8d6f5bce72c2dda34f/crossfit/backend/torch/hf/memory_curve_utils.py#L27-L89
Raised Pytorch issue here: https://github.com/pytorch/pytorch/issues/133280
Another related issue is: https://github.com/pytorch/pytorch/issues/133281
Currently our memory estimation works only with
torch's
default memory allocator, we should make sure it works with RMM allocator plugged in too.Like ensure below works for RMM backed Pytorch allocator too. https://github.com/rapidsai/crossfit/blob/6c213c96b3366298c5681c8d6f5bce72c2dda34f/crossfit/backend/torch/hf/memory_curve_utils.py#L27-L89