The train code not compatible with latest accelerate 1.8.1, please fix it to 1.6.0

[ILG2021]

Checks

• This template is only for bug reports, usage problems go with 'Help Wanted'.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

not related

Steps to Reproduce

When the train goes to validate and save checkpoint it will crash by:
self.accelerator.unwrap_model(self.optimizer).state_dict()
it says that self.optimizer is not a nn.Module class

✔️ Expected Behavior

No response

❌ Actual Behavior

No response

[SWivid]

cc @ZhikangNiu , have you encountered this recently with high version hf accelerate?

@ILG2021 are you using bnb 8bit adam?

F5-TTS/src/f5_tts/model/trainer.py

Line 138 in ebbd7bd

self.optimizer = bnb.optim.AdamW8bit(model.parameters(), lr=learning_rate)

would be good if full error log provided

[ZhikangNiu]

cc @ZhikangNiu , have you encountered this recently with high version hf accelerate?

@ILG2021 are you using bnb 8bit adam?

F5-TTS/src/f5_tts/model/trainer.py

Line 138 in ebbd7bd

self.optimizer = bnb.optim.AdamW8bit(model.parameters(), lr=learning_rate)

would be good if full error log provided

haven't try accelerate 1.8.1

[ZhikangNiu]

@ILG2021 How did you solve it?

[ILG2021]

@ILG2021 How did you solve it?

I reinstall 1.6.0. Everything is ok now.

[ZhikangNiu]

@SWivid peining chen also encountered this problem, and reinstall accelerate 1.6.0 can solve this problem. But when I try torch 2.7 and accelerate=1.8.1, it can work well.