Ross Wightman

@rsomani95 the components themselves are equivalent at a functional level, but the naming was remapped, so would have to remap for this model as well...

@rsomani95 I took a closer look at this s1/s2 (mc1/mc2) are the easiest, could probably map those to OpenCLIP w/ a timm FastViT encoder (after a few additions and a...

@rsomani95 yup, I was co-ordinating with them to set it up. timm and OpenCLIP are already pointing at those checkpoints.

also worth pointing out, timm is supporting all of the models, incl s0 as it's image-tower only. OpenCLIP isn't supporting S0 because it is too much extra work to support...

@MengqingCao see #2138 ... I need a better design to centralize device specific acccelerator module loading, etc instead of spreading it out across many files, it's not a sustainable approach....

@MengqingCao I don't really have any way to test this so don't want to have support for other hardware like this touching as many files. Same thing for Intel and...

@cfgfung Hi, thanks for the PR, I have to come up with a better plan for handling the growing number of accel differences, with #2109 etc need a better way...

uniqueness doesn't really matter here... we and CLIP_benchmark are using the same class names and prompts as OpenAI did for their original CLIP evals. Changing that would alter the results...

@NMontanaBrown hmm, that's not a direct dependency of this project, it's downstream dependency of datasets and would only be used if using a dataset via HF `datasets`, so seems like...

will make note though, perhaps add something to the README, there are also some gotchas related to datasets in terms of fsspec versions that don't appear to be fully spelled...