transformers

bb666bd0 - FP8 backward: chunked expert dequant, skip autograd.Function at inference (#50117)

FP8 backward: chunked expert dequant, skip autograd.Function at inference (#50117) * Pass the input gradient through FP8 layers, so adapters can train on top of them * Same backward for FP8GroupedLinear, with a contiguous gradient for the grouped matmul * Dequantize in place in the FP8 backward * Fit the EP sentinel comment in five lines * FP8 backward: chunk expert dequant, broadcast scales, skip autograd.Function when no grad is needed * Only chunk the FP8 grouped backward when the expert stack exceeds 1 GiB * Free each dequantized expert chunk before the next one * Update measured peak in comment --------- Co-authored-by: Quentin Gallouédec <quentin.gallouedec@huggingface.co> Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com> Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>