lightseq
lightseq copied to clipboard
about Bart export
in examples/inference/python/export/export_bart.py, why do u do Fix encoder layer {} LayerNorm scale and bias Fix decoder layer {} LayerNorm scale and bias
it seems to assign the weight of layer(i+1) = layer(i).
the Bart model in huggingface have no such ops.
That is a trick to support pre-LayerNorm and post-LayerNorm, we will remove those confusing code in the future.