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.

Sign up for GitHub

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

Initial kernel changes to support GaLore #1137

Open
wants to merge 8 commits into
base: main
Choose a base branch
from
+291 −184
Open

Initial kernel changes to support GaLore #1137

Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
a5f9552
Initial kernel changes for 2-state optimizers to support GaLore
matthewdouglas Mar 18, 2024
57de0d6
Experimental implementation for bnb.optim.GaLoreAdamW8bit
matthewdouglas Mar 19, 2024
759c78d
Cleanup
matthewdouglas Mar 19, 2024
b2a01ec
Fix mistake
matthewdouglas Mar 19, 2024
31854da
One more time
matthewdouglas Mar 19, 2024
b1fb85b
Support eturn_outputs buffer option for 1-state optimizers
matthewdouglas Apr 7, 2024
0c87b70
fix wheel build on windows
matthewdouglas Oct 30, 2024
7e3b5ff
Introducte GaLoreWrappedParameter to decouple grad from param
matthewdouglas Oct 30, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Fix mistake
@matthewdouglas
matthewdouglas committed Oct 30, 2024
commit b2a01ec90fb23e76f27fac6acc4603925e660a05
2 changes: 1 addition & 1 deletion bitsandbytes/optim/adamw.py
View file
Original file line number Diff line number Diff line change
@@ -228,7 +228,7 @@ def step(self, closure=None):
self.prefetch_state(p)

if "rank" in group:
self.update_step(group, p, gindex, pindex, return_updates=lor_update)
self.update_step(group, lor_update, gindex, pindex, return_updates=lor_update)

# GaLore Projection Back
p.data.add_(state["projector"].project_back(lor_update))
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

From Algorithm 1 in the paper:

update = project_back(lor_update)
weight.data += update