Add Beam Search sampler #618
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
rlouf
force-pushed
the
beam-search-implementation
branch
10 times, most recently
from
137740a
to
a06f4b5
Compare
rlouf
force-pushed
the
beam-search-implementation
branch
from
16dc89e
to
f4c1eeb
Compare
lapp0
reviewed
Feb 8, 2024
docs/reference/samplers.md
Outdated
| from outlines import models, generate, samplers | ||
|
|
||
|
|
||
| model = models.transformers("mistralai/Mistral-7B-0.1") |
There was a problem hiding this comment.
Typo, it should be mistralai/Mistral-7B-v0.1
However I think we should be recommending mistralai/Mistral-7B-Instruct-v0.2
|
I overlooked something: some Beam Search implementations clone each beam K times and then down sample them to preserve some kind of sample diversity, see this implementation for instance. This can be done in another PR. |
rlouf
force-pushed
the
beam-search-implementation
branch
2 times, most recently
from
d23d3fd
to
c269c0b
Compare
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Outdated suggestions cannot be applied.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.
Closes #258.
In order to implement Beam Search I had to make a few changes to the samplers, they now:
token_ids's ancestors, i.e. the sequence to which they need to be added. While trivial for greedy and multinomial sampling, this is what will allow us to update "beams" in beam search.token_ids,attention_masks,kv_cache,fsmandfsm_statesare now updated using theancestorsinformation.GenerationStatecontains the sequence's weights and ancestors so we can inspect the sampling process.I also simplified the
get_generated_token_idsmethod ofSequenceGenerator. We should soon do a cleaning pass on this class: remove deprecated init arguments, make some methods independant functions and test them.