Enable the combination of sequence length warmup and RoPE #285
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
conglongli
requested review from
jeffra,
tjruwase,
ShadenSmith,
awan-10,
eltonzheng,
minjiaz,
RezaYazdaniAminabadi,
duli2012,
mrwyattii,
arashb,
xiaoxiawu-microsoft and
GuanhuaWang
as code owners
conglongli
requested review from
tohtana
and removed request for
arashb,
ShadenSmith,
jeffra,
duli2012,
tjruwase,
awan-10,
GuanhuaWang,
mrwyattii,
eltonzheng,
minjiaz,
RezaYazdaniAminabadi and
xiaoxiawu-microsoft
tohtana
approved these changes
Nov 9, 2023
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Outdated suggestions cannot be applied.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.
Enable the combination of sequence length warmup (SLW, one of the curriculum learning techniques released by the DeepSpeed team: https://www.deepspeed.ai/tutorials/curriculum-learning/ and https://www.deepspeed.ai/tutorials/data-efficiency/) and RoPE (rotary positional embedding).
Verified the effectiveness of the sequence length warmup when RoPE is used together:
Blue line: baseline with RoPE and without SLW. GPT-3 1.3B pretraining with 30B tokens. Used learning rate = 1.0e-3 (5x of OpenAI's original recipe) and batch size = 2048 (4x of OpenAI's original recipe) to simulate an unstable training.
Red line: with RoPE and with SLW (cl_min = 64, cl_step = 3000). GPT-3 1.3B pretraining with 30B tokens. Same learning rate and batch size as blue line. Compared to blue line, avoided bigger loss spike and achieved better validation loss at the end of training.