Mode 'parallel' for EncSALayer to speed up infer on ONNX #191
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
'transformer-parallel' is widely used in GPT-J-6B and has been proven to have the same effect as traditional transformer.
It can be simplified as:
This saves a skip link and a LayerNorm.This can bring a slight improvement in training speed on Diffsinger.
After experimentation, this modification has shown a more significant improvement on ONNX.
The following are the experimental parameters and results. The benchmark was performed using
infer_acoustic.py
, and the backbone of the model used lynxnet, without using shallow diffusion.On average, the inference speed has increased by 8%.
This change has been applied to yousaV1.42ReFlow and there have been no reports of any issues yet.