[Speculative decoding 2/9] Multi-step worker#1
Closed
cadedaniel wants to merge 30 commits intorejection-samplerfrom
Closed
[Speculative decoding 2/9] Multi-step worker#1cadedaniel wants to merge 30 commits intorejection-samplerfrom
cadedaniel wants to merge 30 commits intorejection-samplerfrom
Conversation
cadedaniel
changed the title
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
cadedaniel
force-pushed
the
multi-step-worker
branch
2 times, most recently
from
391f76a to
4269c84
Compare
cadedaniel
force-pushed
the
multi-step-worker
branch
from
4269c84 to
b20ed29
Compare
* Align top_p and top_k with huggingface * remove _get_prompt_and_output_tokens * rename _apply_top_p_top_k * compare top_p top_k with hf * fix test errors
…m-multi-step-worker-pr
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
18 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Speculative decoding requires running a draft model several times. If this is orchestrated by the LLMEngine process, we incur high overhead going back and forth to the workers which hold the draft model.
To fix this, we instead allow a worker to execute multiple forward passes.