-
-
Notifications
You must be signed in to change notification settings - Fork 5.9k
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
[V1] Add logits_processors support #12688
Conversation
Signed-off-by: imkero <kerorek@outlook.com>
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
Signed-off-by: imkero <kerorek@outlook.com>
Signed-off-by: imkero <kerorek@outlook.com>
Signed-off-by: imkero <kerorek@outlook.com>
Signed-off-by: imkero <kerorek@outlook.com>
Signed-off-by: imkero <kerorek@outlook.com>
Maybe still some overhead near parameters = inspect.signature(logits_processor).parameters trying to enhance it |
Signed-off-by: imkero <kerorek@outlook.com>
IMO we should take the V1 switch as an opportunity to change the design for logits processors. In particular I don't think we should support passing arbitrary LPs themselves per request (in Also I think we agreed that the interface should change to be stateful and work at a batch level (aligned with V1 persistent batch) so that vectorized implementations will be possible/encouraged by default. |
+1 for @njhill's comment |
Add
SamplingParams.logits_processors
support for V1