merge: Pull Huihuo's data fix into microsoft-main
#63
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
saforem2
commented
Oct 12, 2024
megatron/data/gpt_dataset.py
Outdated
Comment on lines
119
to
121
| np_rng = np.random.RandomState(seed=dataset_builders[0].seed) | ||
| self.shuffle_index = np.arange(self.num_samples) | ||
| np_rng.shuffle(self.shuffle_index) |
There was a problem hiding this comment.
Create (and shuffle) a global shuffle_index of len(num_samples).
Explicitly, this
BuildConcatDataset.shuffle_indexis a list of indices maps each sample to a particular {dataset_index, dataset_sample_index}
Example
>>> import numpy as np
>>> shuffle_index = np.arange(10)
>>> np_rng = np.random.RandomState(seed=123)
>>> shuffle_index
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np_rng.shuffle(shuffle_index)
>>> shuffle_index
array([4, 0, 7, 5, 8, 3, 1, 6, 9, 2])…pSpeed into feature/profile
Other quality of life improvementsNote Below copied from Logging Improvements
Changes to
|
…into hzheng-data-fix
…tron-DeepSpeed into hzheng-data-fix
Pull in changes from [6acc370](6acc370) to [`megatron/utils.py`](https://github.com/argonne-lcf/Megatron-DeepSpeed)
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Outdated suggestions cannot be applied.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.
Data Fix
Adds mechanism for correctly shuffling samples across documents from multiple corpora.
The mechanism is implemented inside the
BuildConcatDatasetobject frommegatron/data/gpt_dataset.py.In particular, the shuffle indices are created at
gpt_dataset.py#L118-119which are then used for selecting an individual sample in the
BuildConcatDataset.__getitem__method here:gpt_dataset.py#L130-L132Other changes
ALCF/requirements/requirements.txtALCF/test_blendable_dataset.pymegatron/core/pipeline_parallel/p2p_communication.pymegatron/data/gpt_dataset.pymegatron/utils.py