-
Notifications
You must be signed in to change notification settings - Fork 216
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add UL2 data sampling and pretraining #358
base: main
Are you sure you want to change the base?
Conversation
Since we create them in the T5 data loader, why not use them?
Handles backward-compatibility, so the rest of the code base does not need to change.
Namely sampling from uniform and normal distributions.
db95ce8
to
4d9ff77
Compare
... which also improve error messages.
Instead, the user should choose a larger maximum sequence length, which an error warns them about.
Previously, I truncated sequences so the maximum amount of duplicated This is probably a worse/undesired solution, so I kept the other code in for now (but commented). Note that erroring out is also how the |
Instead of concatenating arrays and lists to get a certain dtype.
megatron/data/dataset_utils.py
Outdated
# Note(mingdachen): | ||
# By default, we set the probilities to favor shorter ngram sequences. | ||
pvals = 1. / np.arange(1, max_ngrams + 1) | ||
pvals /= pvals.sum(keepdims=True) | ||
if favor_longer_ngram: | ||
pvals = pvals[::-1] | ||
elif sampling_style is SamplingStyle.NORMAL: | ||
normal_mean = (max_ngrams + 1) / 2 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
normal_mean
is not used it seems
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For small sequence lengths or low probability/mean ngram values, we could get `max_ngrams` < 1 and `max_predictions_per_seq` < 1, causing no masking to be done.
Now same as in the UL2 paper code snippet.
There were several issues still remaining in the UL2 implementation, most notably that I only tested for micro batch sizes of 1, which when increased made the decoder-only models fail. :p The implementation also more closely follows the |
As in the T5 codebase. This could have highly detrimental effects on performance of TorchScript cannot easily type-dispatch the `bias_dropout_add` function.
More code reuse, change some methods to functions and change their visibility.
For readability.
By pre-allocating more data.
I can finally report results... Comparing standard T5 training vs training with UL2 or UL2R, results in lm-eval-harness were almost always better with UL2/UL2R. Which should mean this code does improve evaluation results. :) |
DS = DeepSpeed No idea why this happens, I couldn't explain it after briefly looking into the DeepSpeed source.
That is, the reproduced objective token.
Was missing `max_seq_length_dec`.
This was already the case for encoder-decoders, but is now also the case for decoder-only models.
This also fixes problems with decoder-only attention masks.
When using the custom fused softmax kernel.
This adds pretraining using UL2 for both encoder-decoder, non-causal decoder-only, and causal decoder-only models.
I have not yet run large-scale tests to see if it yields the desired training improvements, but I wanted to give others the option to take a look at the code already.