feat(server): reduce memory requirement #214
Merged
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
|
@njhill, tests are failing but I like the idea. |
njhill
changed the title
|
@OlivierDehaene thanks, let me fix the tests soon. |
njhill
force-pushed
the
tgi/mem_reduction
branch
from
e97abd2
to
0b1d001
Compare
|
@OlivierDehaene tests fixed and I've also added the same for seq2seq_lm. |
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Outdated suggestions cannot be applied.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.
Currently the memory requirement is much higher than it needs to be due the way that past key values are pruned and concatenated.
This PR changes
CausalLMto do these operations layer by layer so that the corresponding tensors can be freed incrementally. It should roughly halve the memory requirement and so allow much larger batch sizes and/or sequence lengths.Also changed the pruning to remove any unused padding at the front the batch (for example when the longest input sequence in the batch gets pruned).
These changes are also kind of required for the non-flash-attention impl of #210 (which I can push soon), to ensure that assumptions around mem usage based on point-in-time batch contents remain valid.
@OlivierDehaene if you approve of this I can add the corresponding
Seq2SeqLMchanges.