Matt
Matt
Hi @jenkspt, that Mixtral PR I made is a little out of date, and will be updated soon. The API we're intending now is quite close to the OpenAI API:...
@CISC Not especially! We do have one or two other unavoidable divergences from the OpenAI API. Most notably, we allow a `return` key in the JSON schema for tools. This...
@CISC My bad, actually! The role has already been updated from `tool_results` to `tool` as you suggested, I just forgot that I made that commit :sweat:
I'm not really sure what the solution is, unfortunately - thankfully the issue is mostly limited to `BlenderBot`, but the pipeline can't know if it'll be used for chat or...
@a8nova that happens a lot - it means it's a custom code model. Those models include their modelling source in the repo itself with the weights, which means they don't...
Hi @ZhiyuanChen, I don't think we will ever achieve complete MyPy compatibility, unfortunately! Many of our functions are extremely polymorphic, and even enumerating all the possibilities in the return type...
This should be ready - cc @amyeroberts for core maintainer review! I'm speeding things up because the lack of this feature is causing some user issues, and we have a...
I think this PR is causing issues in the CI because `pt_cuda_available` defaults to "NA", which is a truthy value, and so the `if pt_cuda_available:` block is executed even if...
Opened a fix at #31113!
Hey @a8nova, I'm sorry I lost track of this! There's actually been some changes - we're currently trying to deprioritize adding new TF models to Transformers, in favour of improving...