[Storage Cleaner] Speed up unsharding of some legacy checkpoints #488
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
epwalsh
reviewed
Mar 7, 2024
olmo/checkpoint.py
Outdated
| if rank_size == 0: | ||
| return | ||
|
|
||
| temp: np.ndarray = torch.zeros(rank_size, dtype=shard0_md.tensor_properties.dtype).numpy() |
There was a problem hiding this comment.
What's the purpose of this temp array? If it's just to get the number of bytes, you should be able to infer that from the data type and size.
There was a problem hiding this comment.
Just type and number of bytes. I've changed the code to assume fp32 (c58f4b4). I already know they are not bf16 at least because .numpy() fails for bf16.
There was a problem hiding this comment.
yea when we train in bf16 (or fp16), the main copy of model weights is always fp32
olmo/checkpoint.py
Outdated
Comment on lines
1035
to
1036
| temp: np.ndarray = torch.zeros(1, dtype=shard0_md.tensor_properties.dtype).numpy() | ||
| numpy_type = temp.dtype |
There was a problem hiding this comment.
Similar question here, but looks like it's just to get the data type? It's probably reasonable to assume FP32.
There was a problem hiding this comment.
c58f4b4 Changed to assume fp32
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Outdated suggestions cannot be applied.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.
This PR changes the unsharding mechanism of legacy checkpoints to use processes and shared memory instead of threads. In one case where the world size was 1024, this implementation brought the unsharding time down from 6 hours to 30 minutes. This implementation is slower than the old one at smaller scales, but that is ok.
An option was to keep the old mechanism around in code too, but since we are trying to get rid of legacy sharded checkpoints it doesn't seem worth to keep that code around.