fix: correct dictionary unpacking in recursively_apply function

[qgallouedec]

from transformers.modeling_outputs import CausalLMOutputWithPast
from accelerate.utils import convert_to_fp32
import torch

lm_output = CausalLMOutputWithPast(logits=torch.ones(2, 3, 4))
print(convert_to_fp32(lm_output))

before:

CausalLMOutputWithPast(loss={'logits': tensor([[[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]],

        [[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]]])}, logits=tensor([[[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]],

        [[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]]]), past_key_values=None, hidden_states=None, attentions=None)

after

CausalLMOutputWithPast(loss=None, logits=tensor([[[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]],

        [[1., 1., 1., 1.],
         [1., 1., 1., 1.],
         [1., 1., 1., 1.]]]), past_key_values=None, hidden_states=None, attentions=None)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

Indeed, thanks for the fix !