"Optimistic" is_in generation for openai. #378
Conversation
| except IndexError: | ||
| pass | ||
|
|
||
| greedy = False # we try to generate the full response at each iteration |
There was a problem hiding this comment.
We might as well remove the greedy case altogether
There was a problem hiding this comment.
Say you have two tokenized choices [T1, TX], [T2, TX]. You would unmask the three tokens and request 2 tokens from the API. If the system is very biased towards TX, it may answer [TX, TX], and you have gained nothing. Repeating the same call you would enter in a loop. Thus, it makes sense to run a greedy step next with just T1 and T2 unmasked, such that you are guaranteed to move one token forward.
|
It might take me until Monday to be able to properly review the PR, thank you for contributing! |
|
I pushed changes such that it now computes the prefix matches as sequences of tokens (list of ints) and not as strings. That makes the algorithm 100% consistent with the output generated by the LLM. The code may require some refactoring for readability/testability, but as for the algorithmic solution, I think it is complete. I won't push anything else. |
The current approach is greedy, in the sense that it generates a single token at each steps, asking the API to only generate valid next tokens. This mean having to pay for the prompt tokens for every token generated. This commit takes a more optimistic approach. It starts with allowing all tokens present in the sequences, and limiting the length of the generation to the number of tokens in the longest sequence. If the completion is not satisfactory it then takes one greedy step before switching back to the optimistic mode. On average this new approach consumes less tokens than the current one.
rlouf
force-pushed
the
feature/optimistic_is_in
branch
from
f70570c to
fb1061a
Compare
|
This is a very good addition, thank you! I just rebased the commits into a single commit. I will refactor the entire OpenAI integration soon, so I will leave the code as is for now and merge. |
Addresses #368.
It is quite verbose: adds two helper functions:
longest_common_prefixandget_choices_with_longest_common_prefix.Maybe you would order the code differently.
ITo make the solution 100% secure, instead of computing the "prefixes" string-based, it should be done "token-based". That would change just those two functions. Let me know what you think.
The algorithm follows the pseudo-code in the issue.