Closed uahmed93 closed 3 months ago
in crossfit/backend/torch/loader.py when padded tokens are other than 0 , seq_length will always be max value.
crossfit/backend/torch/loader.py
seq_length = (frame[sort_key] != 0).sum(axis=1)
because of which clipping logic is failing.
Fixed by: https://github.com/rapidsai/crossfit/pull/66
in
crossfit/backend/torch/loader.py
when padded tokens are other than 0 , seq_length will always be max value.because of which clipping logic is failing.