Skip to content

Conversation

@yorickvP
Copy link
Contributor

@yorickvP yorickvP commented Jul 4, 2024

Open questions:

  • Should we keep detokenizing in predict.py instead of postprocessing?
  • It looks like the required config changed. Notably, enable_trt_overlap disappeared and a whole bunch of cache options got added.
  • Yorick: do some testing.

@yorickvP yorickvP requested a review from joehoover July 4, 2024 13:14
@yorickvP yorickvP changed the title tensorrt-llm: 0.9 -> 0.19, triton: 2.42.0 -> 2.44.0 tensorrt-llm: 0.9 -> 0.10, triton: 2.42.0 -> 2.44.0 Jul 5, 2024
@yorickvP yorickvP marked this pull request as ready for review July 10, 2024 15:08
@technillogue
Copy link
Contributor

enable_trt_overlap is set to false in a lot of places, we will probably need to change that

we should review the new configuration options as well

Comment on lines +375 to +377
if token == []:
continue

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

was this discovered by testing? what is this?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good question, it was discovered by testing. EOS seems to have been replaced by []. Potentially related to tokenizer config?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants