Support 4bit BNB layers meta-device materialization #19150
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
carmocca
added
feature
github-actions
bot
added
fabric
carmocca
commented
Dec 13, 2023
carmocca
force-pushed
the
carmocca/bnb-metad
branch
from
d341df3 to
f5be300
Compare
carmocca
force-pushed
the
carmocca/bnb-metad
branch
from
1ec5916 to
ef17e7a
Compare
carmocca
requested review from
awaelchli,
justusschock,
williamFalcon,
lantiga and
Borda
as code owners
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## master #19150 +/- ##
==========================================
- Coverage 83% 54% -29%
==========================================
Files 443 439 -4
Lines 36859 36984 +125
==========================================
- Hits 30539 19940 -10599
- Misses 6320 17044 +10724 |
awaelchli
approved these changes
Dec 20, 2023
Borda
approved these changes
Dec 20, 2023
carmocca
commented
Dec 20, 2023
| @@ -37,7 +39,8 @@ | |||
|
|
|||
| log = logging.getLogger(__name__) | |||
|
|
|||
| _BITSANDBYTES_AVAILABLE = RequirementCache("bitsandbytes>=0.41.0") | |||
| # TODO: unpin after resolving the `quant_state` format breaking changes | |||
| _BITSANDBYTES_AVAILABLE = RequirementCache("bitsandbytes==0.41.0") | |||
There was a problem hiding this comment.
Still new to this repo, so maybe I misunderstood something.
Not sure how it can work.
In requirements/pytorch/extra.txt the BNB is fixed to 0.41.1.
The same goes to tests: skipif will always return True. That's why these tests are skipped.
These are GPU tests
There was a problem hiding this comment.
Oh great catch. I messed upthe requirements, I'll open a PR
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What does this PR do?
Adds support for converting Linear layers on meta-device to Bitsandbytes Linear layers.
Also for materializing and quantizing Bitsandbytes Linear layers on meta-device.
Scenario 1
Cons:
to_emptyandreset_parameters(called frommaterialize_meta_tensors) will create empty tensors and quantize. There's no way to avoid this unless we makematerialize_meta_tensorsaware of bitsandbytes.load_checkpointwill again quantize weights that were already quantized during materialization.Scenario 2
I believe this is the ideal scenario.
Scenario 3
Cons:
fabric.setupwill need to recreate layers so the model hooks are lost.to_emptyandreset_parameters(called frommaterialize_meta_tensors) will both create empty tensors and quantize. There's no way to avoid this unless we makematerialize_meta_tensorsaware of bitsandbytes.Scenario 4
Cons:
8-bit layer materialization is not implemented. I only made the minimal changes required for it.
📚 Documentation preview 📚: https://pytorch-lightning--19150.org.readthedocs.build/en/19150/
cc @Borda @carmocca @justusschock @awaelchli