Skip to content

[autochunk] support multi outputs chunk search #2538

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Feb 1, 2023
Merged

Conversation

oahzxl
Copy link
Contributor

@oahzxl oahzxl commented Feb 1, 2023

Support multi outputs chunk search. Previously we only support single output chunk search. It is more flexible and improve performance by a large margin. For transformer, we reduce memory by 40% than previous search strategy.

  1. rewrite search strategy to support multi outputs chunk search
  2. fix many, many bugs
  3. update tests

* reorder alphafold tests

* unet

* add unet test

* update gpc

* add autochunk available

* update test utils

* update test utils

* code style

* rename

* update evoformer stack

* simplify test

* delete simple evoformer

* update chunk result assert

* update chunk assert

* update diffuser test

* skip duffuser test for now

* init gpt

* refactor op detect

* refact pattern recog

* update some op, fix view bug

* update node

* fix bug

* fix bug

* remove useless part

* finish tf

* code style

* code style

* update chunk select

* support view list, fix many bugs

* finish gpt test

* update gpt test

* fix a bug

* refact some function, clearer

* update doc

* move a function

* seperate find chunk info function

* can run now

* fix view bug

* update memory estimate for multi out

* support split in codegen

* run multi out for transformer

* add print memory

* seperate node manager

* finish gpt test

* update alphafold assert

* update tests

* add benchmark

* fix a bug

* update benchmark

* merge

* remove benchmark
@feifeibear
Copy link
Contributor

How is the speed?

@oahzxl
Copy link
Contributor Author

oahzxl commented Feb 1, 2023

How is the speed?

The speed loses about 3-4%. I will update a benchmark in next pr.

@oahzxl oahzxl merged commit 05671fc into hpcaitech:main Feb 1, 2023
@oahzxl oahzxl deleted the out branch February 2, 2023 07:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants