[llm] Create separate Predictor for LLMs and enable flash attention on CUDA #3409
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
Unit Test Results 6 files ± 0 6 suites ±0 1h 28m 20s ⏱️ + 16m 55s For more details on these failures, see this check. Results for commit bd0ddc2. ± Comparison against base commit cb37535. ♻️ This comment has been updated with latest results. |
for more information, see https://pre-commit.ci
…nto llm-predictor
for more information, see https://pre-commit.ci
…nto llm-predictor
tgaddair
changed the title
arnavgarg1
reviewed
May 21, 2023
arnavgarg1
reviewed
May 21, 2023
arnavgarg1
reviewed
May 21, 2023
Comment on lines
+40
to
+45
| class DictWrapper: | ||
| """Wrapper for a LudwigFeatureDict module that allows for iteration over keys. | ||
|
|
||
| The purpose of this class is to avoid exposing input and output features as modules of the LLM. This is because we | ||
| only wish to train the underlying model, and having these additional modules can confuse systems like DeepSpeed. | ||
| """ |
There was a problem hiding this comment.
Very clever way of getting around initializing all of the other modules
arnavgarg1
reviewed
May 21, 2023
arnavgarg1
reviewed
May 21, 2023
…nto llm-predictor
arnavgarg1
changed the title
|
Failed test looks like a transient issue, it should be safe to merge |
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Outdated suggestions cannot be applied.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.
This PR introduces a new
LlmPredictorclass used only during batch prediction for the purpose of running text generation instead of simply outputting logits (as during training) from the forward pass. This is because at predict time we want the fully generated sequence, not just the logits.Other fixes included in this PR:
LLMmodule.This PR also disables
prompt_tuningas an adapter type for now, as generation mode does not currently work when this adapter is applied (seetests/integration_tests/test_llm.py::test_llm_finetuning_strategies[prompt_tuning_init_random-local]).