(3/n) Support 2D Parallelism - Efficient loading of full-state checkpoints #19870
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
awaelchli
force-pushed
the
feature/tp-full-ckpt-load
branch
from
86d8d16 to
1170618
Compare
awaelchli
force-pushed
the
feature/tp-full-ckpt-load
branch
from
1170618 to
1a0887e
Compare
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
awaelchli
changed the title
awaelchli
commented
May 15, 2024
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #19870 +/- ##
=========================================
- Coverage 84% 59% -25%
=========================================
Files 425 420 -5
Lines 35010 34925 -85
=========================================
- Hits 29369 20527 -8842
- Misses 5641 14398 +8757 |
justusschock
approved these changes
May 15, 2024
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Outdated suggestions cannot be applied.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?
Follow-up to #19852.
I found that loading large full-state-dict checkpoints into a distributed model can lead to OOM (e.g. Llama 3 70B) because PyTorch's approach of loading on rank-0, then broadcasting and redistributing is applied to the entire checkpoint at once, instead of on a per-parameter or per-module basis (see comment).
In this PR, I load the checkpoint per-parameter, which seems to work as it should.
Mini inference benchmark on Llama 3 8B (8xA100)
Before:
4.64 GB (peak memory usage), 13.43 seconds to load
Now:
3.20 GB (peak memory usage), 13.72 seconds to load
Llama 3 70B (8xA100)
Before:
OOM
Now:
20.01 GB (peak memory usage), 40.73 seconds to load
Benchmarks done with this LitGPT branch.
cc @justusschock @awaelchli @carmocca @Borda