-
Notifications
You must be signed in to change notification settings - Fork 3.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Frontend] Add bad_words_ids sampling parameter #5986
base: main
Are you sure you want to change the base?
Conversation
@@ -28,6 +28,77 @@ class SamplingType(IntEnum): | |||
to sample from.""" | |||
|
|||
|
|||
class NoBadWordsLogitsProcessor: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I am not sure if this is the right file for the class. Still I thought it could be placed here near the LogitsProcessor type (the one just above). But if there is a better place for the processor class, I will be ready to move it)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes I think that if we include this it should go in a different file.
if len(bad_word_ids) == 1: # 1-token words already processed | ||
continue | ||
|
||
if len(bad_word_ids) > len(past_tokens_ids) + 1: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This differs from the original inequality here: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L1131
if len(sequence_ids) > input_ids.shape[1]: # the sequence is longer than the context, ignore
continue
I may be wrong, but it seemed we should allow for one token to be generated (so + 1).
I'm unsure whether or not it makes sense to support this. I know it's an option in transfomers but it was added very early on and the implementation seems limited/clunky to me. Wouldn't it make more sense for the bad words to be a list of strings rather than token sequences? |
Yes, I absolutely agree that this "list of lists of token ids" structure is not very friendly and easy to use 😅 Sure, list of strings will be more convenient. I believe I can change that. Just wanted to clarify some questions before actually making some changes. If we make bad words as list of strings (and call it, for example, just So, the main and only question is actually the following — should we keep transformers' "clumsy" Oh, and one more point. Currently, SamplingParams has Hmm, just noticed, SamplingParams actually has another parameter — Ok, I agree that introducing |
My view on this kind of thing is to collect some concrete requirements / use cases and base on that. I.e. avoid adding things with hypothetical benefit. Would be good to see some explicit examples of how/where this functionality is used, and that should then also inform what kind of thing makes the most sense w.r.t. the various options being discussed. |
Bad words ids -> Bad words
I looked through some examples and use cases. First, it seems that However, seems like it would be indeed better to have a list of words and not token sequences. Because people always have to use tokenizers in order to get token ids. This is just stated in the documentation on the transformers docs site. Usually, these tokenizer things lead to questions (not very friendly "low level" functionality). I also found a few mentions of bad words list (not token ids) functionality in other repositories. For example, in TensorRT-LLM Backend, there is already an option to provide a list of bad words as strings. In outlines, there are plans to add this feature. So, I updated the PR: now there is However, to some extent, this complicated the implementation. So that I am currently not 100% sure that it is just a "Frontend" related feature 😅 @njhill could you please take a look and share your thoughts on the updated code? Points which are to be resolved (if the general idea would seem OK, I will fix this):
P.S. For the history, this is the PR which added |
Forgot to add, in vLLM, there is already something like bad words ids thing: logit_bias_logits_processor, which is added for the compatibility with OpenAI request params. |
FIX #986
PR Checklist (Click to Expand)
Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.
PR Title and Classification
Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:
[Bugfix]
for bug fixes.[CI/Build]
for build or continuous integration improvements.[Doc]
for documentation fixes and improvements.[Model]
for adding a new model or improving an existing model. Model name should appear in the title.[Frontend]
For changes on the vLLM frontend (e.g., OpenAI API server,LLM
class, etc.)[Kernel]
for changes affecting CUDA kernels or other compute kernels.[Core]
for changes in the core vLLM logic (e.g.,LLMEngine
,AsyncLLMEngine
,Scheduler
, etc.)[Hardware][Vendor]
for hardware-specific changes. Vendor name should appear in the prefix (e.g.,[Hardware][AMD]
).[Misc]
for PRs that do not fit the above categories. Please use this sparingly.Note: If the PR spans more than one category, please include all relevant prefixes.
Code Quality
The PR need to meet the following code quality standards:
format.sh
to format your code.docs/source/
if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.Notes for Large Changes
Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with
rfc-required
and might not go through the PR.What to Expect for the Reviews
The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:
action-required
label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.Thank You
Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!