fix last tokens passing to sample_repetition_penalties function#1295
Merged
abetlen merged 2 commits intoabetlen:mainfrom Apr 1, 2024
Merged
fix last tokens passing to sample_repetition_penalties function#1295abetlen merged 2 commits intoabetlen:mainfrom
abetlen merged 2 commits intoabetlen:mainfrom
Conversation
ymikhailov
force-pushed
the
fix-last-tokens-passing
branch
from
3e06ff1 to
7f83e1d
Compare
Owner
|
@ymikhailov thank you for the fix! Looks good. |
xhedit
pushed a commit
to xhedit/llama-cpp-conv
that referenced
this pull request
Apr 6, 2024
…tlen#1295) Co-authored-by: ymikhaylov <ymikhaylov@x5.ru> Co-authored-by: Andrei <abetlen@gmail.com>
xhedit
added a commit
to xhedit/llama-cpp-conv
that referenced
this pull request
Apr 6, 2024
* feat: add support for KV cache quantization options (abetlen#1307) * add KV cache quantization options abetlen#1220 abetlen#1305 * Add ggml_type * Use ggml_type instead of string for quantization * Add server support --------- Co-authored-by: Andrei Betlen <abetlen@gmail.com> * fix: Changed local API doc references to hosted (abetlen#1317) * chore: Bump version * fix: last tokens passing to sample_repetition_penalties function (abetlen#1295) Co-authored-by: ymikhaylov <ymikhaylov@x5.ru> Co-authored-by: Andrei <abetlen@gmail.com> * feat: Update llama.cpp * fix: segfault when logits_all=False. Closes abetlen#1319 * feat: Binary wheels for CPU, CUDA (12.1 - 12.3), Metal (abetlen#1247) * Generate binary wheel index on release * Add total release downloads badge * Update download label * Use official cibuildwheel action * Add workflows to build CUDA and Metal wheels * Update generate index workflow * Update workflow name * feat: Update llama.cpp * chore: Bump version * fix(ci): use correct script name * docs: LLAMA_CUBLAS -> LLAMA_CUDA * docs: Add docs explaining how to install pre-built wheels. * docs: Rename cuBLAS section to CUDA * fix(docs): incorrect tool_choice example (abetlen#1330) * feat: Update llama.cpp * fix: missing logprobs in response, incorrect response type for functionary, minor type issues. Closes abetlen#1328 abetlen#1314 * fix: missing logprobs in response, incorrect response type for functionary, minor type issues. Closes abetlen#1328 Closes abetlen#1314 * feat: Update llama.cpp * fix: Always embed metal library. Closes abetlen#1332 * feat: Update llama.cpp * chore: Bump version --------- Co-authored-by: Limour <93720049+Limour-dev@users.noreply.github.com> Co-authored-by: Andrei Betlen <abetlen@gmail.com> Co-authored-by: lawfordp2017 <lawfordp@gmail.com> Co-authored-by: Yuri Mikhailov <bitsharp@gmail.com> Co-authored-by: ymikhaylov <ymikhaylov@x5.ru> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
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
2 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.
Bug: currently all previous tokens are passing to
llama_sample_repetition_penaltiesfunction instead of lastpenalty_last_ntokens.Description:
llama_sample_repetition_penaltiesfunction expects lastpenalty_last_ntokens. But it gets all previously generated tokens which leads to the algorithm not working correctly.As you can see in the code below it will use only first
penalty_last_ntokens to createtoken_countdict:This behavior is obviously incorrect which in our case led to the case where our model had a lot of answers with repetitions.
Also in this PR was fixed bug with possible buffer overflow that could cause to segfault in code above.