About training 1.2M updates over one week mentioned in the paper

Checks

• This template is only for research question, not usage problems, feature requests or bug reports.
• I have thoroughly reviewed the project documentation and read the related paper(s).
• I have searched for existing issues, including closed ones, no similar questions.
• I am using English to submit this issue to facilitate community communication.

Question details

I have a question about the training updates mentioned in the F5-TTS paper.

According to the paper, the model was pretrained for 1.2M updates, and it was completed in about 1 week.

But with my settings (which follow the paper's configuration):

• batch_size_per_gpu: 38400
• batch_size_type: frame
• max_samples: 64
• grad_accumulation_steps: 4
• 8×GPU training with Accelerate

Each update takes about 4 seconds. In one week, the maximum updates I can get are: 7 × 24 × 3600 / 4 ≈ 150k updates, which is far less than 1.2M updates.

Could you please clarify:

1. Is 1.2M updates the total number of 8xGPU?
2. What is the actual training speed / step time in your environment?
3. How long did it actually take to reach 1.2M updates?

Thank you very much for your help!