A bug in the function convert_to_fp32()

[xyh97]

Hi, recently I found that there might be a bug in the function convert_to_fp32().

accelerate/src/accelerate/utils.py

Line 258 in 1b1463f

return recursively_apply(_is_fp16_tensor, tensor, test_type=_is_fp16_tensor)

I think the correct one should be
return recursively_apply(_convert_to_fp32, tensor, test_type=_is_fp16_tensor)

Thanks.

[reppy4620]

I met the error because of this i think.

If set fp_16 = True in Accelerator, model returns "True".

[sgugger]

Thanks for the report and the hint for the fix. It's fixed by the PR mentioned above and I will do a patch release on Monday with this fix.