Combined dataset feature

[le1nux]

What does this PR do?

This PR addresses issue #258 (inefficiencies in the dataloader) and additionally introduces a combined dataset, where a dataset can now comprise a list of datasets and iterate over them.
As part of fixing the dataloader inefficiencies, we now implement the sample skipping functionality not on the dataloader level anymore but in an adapted version of the PyTorch DistributedSampler. I reran a warm start and the learning is equivalent to a full, non-warmstarted run.

General Changes

• Introduced ResumableDistributedSampler which is a copy of the PyTorch DistributedSampler added with the feature to skip samples. This is from now on used for warmstarts instead of the skip_num_samples in the Dataloader. In case of skipping samples, the dataloader had to instantiate a ResumableBatchSampler which was internally iterating over all the dataset indices. For small datasets this was fine, but for larger datasets (in the trillion token range) this became a bottleneck at instantiation time:
  

  modalities/src/modalities/dataloader/samplers.py

  Lines 25 to 28 in b79d04d

  	self.underlying_batch_sampler = underlying_batch_sampler
  	# NOTE: we are only iterating ove the indices not the actual data
  	# so this is relatively cheap
  	self.indices = list(iter(self.underlying_batch_sampler))

  
  Skipping in the ResumableDistributedSampler is skipping in O(1) now. The ResumableBatchSampler was removed from the codebase.
• Replaced the packed index generation routine (inefficient due to for loop)
  

  modalities/src/modalities/dataloader/dataset.py

  Lines 331 to 334 in b79d04d

  	return [
  	((i * self.block_size - i) * self._token_size_in_bytes, self.block_size * self._token_size_in_bytes)
  	for i in range(num_samples)
  	]

  
  with a vectorized version.
• added new NumberConversion routine num_samples_from_num_tokens

Breaking Changes

• Removed RepeatingDataloader, as a feature that was never actively used for running multiple epochs and had complex maintenance when refactoring the sampling. If needed we could reimpliment it.
• In the settings, the training_progress section has now num_seen_samples instead of local_num_seen_batches , as skipping is now done on the Sampler level and not on the dataloader level anymore
• batch_size and fast_forward_batch_id fields in the LLMDataLoader are not neede anymore and were removed.

Checklist before submitting final PR

• My PR is minimal and addresses one issue in isolation
• I have merged the latest version of the target branch into this feature branch
• I have reviewed my own code w.r.t. correct implementation, missing type hints, proper documentation, etc.
• I have run a sample config for model training
• I have checked that all tests run through (python tests/tests.py)
• I have updated the internal changelog (CHANGELOG_DEV.md)

[flxst]

LGTM :) Left a few minor comments.

[flxst]

Wouldn't it be more consistent if this defaulted to False like in PackedMemMapDatasetContinuous (see line 308)?

[le1nux]

In PackedMemMapDatasetContinuous we would never load the index (apart for debugging purposes), that's why I made it defaulting to False. The "Continuuous" implementation does not need an index. The PackedMemMapDatasetBase, however, in it's default implementation would use the index for packing the data, which is why it defaults to True.

[flxst]

General comments on the changelog:

• The table in the beginning also needs to be updated
• It might be useful to reverse the order of PRs so that the newest ones come first