Conversation
| loader = get_loader(get_train_dataset( | ||
| # Set repeat=False to avoid repeating the dataset. | ||
| # Also add RedistributeLoader to synchronize the end of rank exhaustion. Only works with initialized torch distributed. | ||
| loader = RedistributeLoader(get_loader(get_train_dataset( |
There was a problem hiding this comment.
I think it would be better if this was transparent to the user, so make get_loader handle this internally.
And since the choice of RedistributeLoader vs. StopFirstLoader actually changes the data that's being iterated, this choice should be made in the metadataset and not in the code, I think.
As a property of blend_epochized.
I.e. blend_epochized can either be a list as before (chooses default RedistributeLoader), or it can be a dict for more customization like
blend_epochized:
phase_out_behavior: stop_first_loader
datasets:
- repetitions: 5
path: ./coco
# ... Other parameters
- repetitions: 2
path: ./coyo
- repetitions: 1
path: ./coyo
split_part: val
There was a problem hiding this comment.
Okay, we can make this be handled in get_loader. I personally would prefer it to be separate though, as it's a feature on top?
Regarding moving the configuration to the metadataset:
I see your point that this slightly modifies how the data is iterated, but I'd also argue:
- So far we don't really rely on torch distributed, where this piece of code is tightly bound to it.
- This would also disable nesting of
blend_epochized, because you cannot nest different (or unconfigured)phase_out_behavior. - This depends on
repeat=Falseand doesn't make sense ifrepeat=True, so it's based on what the user sets in the code. - At least for
RedistributeLoaderit should not really change the data frequency (so far the settings in the metadataset mainly focuses on data frequency / blend). - If we move the boundary of metadataset to include this, then we should also have gradient accumulation, seeds, batch size, handling of incomplete batches, etc. in the config. I wouldn't want that, tbh.
Thus voting for keeping this in code, not in the metadataset config.
| except StopIteration: | ||
| # print(f"[r={rank}]: StopIteration\n", end="") | ||
| self.exhausted_states[rank] = self_exhausted = 1 | ||
| dist.all_reduce( |
There was a problem hiding this comment.
We should evaluate the impact of this synchronization (which happens in every iteration!) in a real world training to see if the training speed suffers when we have many nodes and ranks.
There was a problem hiding this comment.
Yes, agreed. Didn't benchmark the impact of this yet.
| return f"RedistributeLoaderState(inner_state={self.inner_state!r}, exhausted_state={self.exhausted_state!r}, overuse_count={self.overuse_count!r})" | ||
|
|
||
|
|
||
| class RedistributeLoader(Generic[T]): |
There was a problem hiding this comment.
I think this will break our reproducible scaling!
It seems the whole redistribution loader will not work if we stop and resume with a different number of ranks.
See #80
That's a problem we need to discuss possible solutions for.
There was a problem hiding this comment.
Okay, haven't thought of this tbh. Yes, let's discuss offline
| return f"StopFirstDataLoaderState(inner_state={self.inner_state!r}, iterating_from_start={self.iterating_from_start!r}, next_sample_restore_key={self.next_sample_restore_key!r})" | ||
|
|
||
|
|
||
| class StopFirstLoader(Generic[T]): |
There was a problem hiding this comment.
Same issue with reproducible scaling here? I think this one could be easier to get it working without breaking repro scaling.
There was a problem hiding this comment.
Likely easier, but it's also the less useful of the two methods 😅
2 participants
Provides two strategies for synchronizing the end of the loader for repeat=False.