Skip to content
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

feat: move transposes up (and then down) #1099

Merged
merged 8 commits into from
Apr 1, 2025
Merged

feat: move transposes up (and then down) #1099

merged 8 commits into from
Apr 1, 2025

Conversation

avik-pal
Copy link
Collaborator

No description provided.

@avik-pal
Copy link
Collaborator Author

forgot to add broadcastindim is reshape to td, though I will merge it with the canon today anyways

@avik-pal avik-pal force-pushed the ap/opts_new branch 4 times, most recently from d409a35 to 2297685 Compare March 31, 2025 18:06
@avik-pal
Copy link
Collaborator Author

i might have enabled conflicting optimizations

@avik-pal avik-pal force-pushed the ap/opts_new branch 3 times, most recently from 632e1ca to 7518034 Compare March 31, 2025 23:46
@avik-pal avik-pal changed the title feat: move transposes up feat: move transposes up (and then down) Mar 31, 2025
@avik-pal avik-pal force-pushed the ap/opts_new branch 2 times, most recently from ad689d8 to 169b182 Compare April 1, 2025 00:23
@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

failures due to EnzymeAD/Enzyme-JAX#586

@giordano giordano force-pushed the ap/opts_new branch 2 times, most recently from 55d215b to 38e713d Compare April 1, 2025 10:40
@giordano
Copy link
Member

giordano commented Apr 1, 2025

This is not working in GB-25: PRONTOLab/GB-25#128 (comment)

@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

Let me try and patch it

@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

and I broke more stuff...

@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

Also @wsmoses one of the 5 commented out reshape passes are probably broken https://github.com/EnzymeAD/Reactant.jl/actions/runs/14195378107/job/39769233267?pr=1099#step:10:576

@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

actually need to check which one is broken, that is not the one I activated in the last commit

@avik-pal
Copy link
Collaborator Author

avik-pal commented Apr 1, 2025

Sp the problem in the GB repo is not fixed yet. The thing which ends up happening is that the input buffers can now be aliased and we need to be careful while donating it

@avik-pal avik-pal force-pushed the ap/opts_new branch 4 times, most recently from dd8a43b to 1e01733 Compare April 1, 2025 16:24
avik-pal added 3 commits April 1, 2025 14:34
feat: try collapsing the ops again

fix: syntax

feat: bump jll

fix: print for debugging

fix: unary transpose

fix: try fixing

fix: conflicting passes

fix: conflicting passes

feat: newer opts

fix: reshape pad

fix: avoid transpose_reduce_window for now

Update to Reactant_jll v0.0.116

fix: add some aliasing checks

feat: enable more passes

fix: disable some passes again

fix: types

fix: rollback some changes

fix: rollback more passes

feat: store if an arg is donated

fix: mark donated

fix: remove broadcast_in_dim_is_reshape for now

fix: reshape_pad is also broken

fix: concat is busted not pad
@avik-pal avik-pal merged commit fd1037a into main Apr 1, 2025
54 of 56 checks passed
@avik-pal avik-pal deleted the ap/opts_new branch April 1, 2025 19:46
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