I think if you read the issue carefully you would understand that the CLIP implementation in transformers and as published by OpenAI is wrong and does not match

I think if you read the issue carefully you would understand that the CLIP implementation in transformers and as published by OpenAI is wrong and does not match their trained model code; and that doing the fix I suggest, empirically for me and in theory, improves results.