Skip to content

fix fft dtype compatibility in deepspeed #1427

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Artiprocher merged 1 commit into from
May 6, 2026
Merged

fix fft dtype compatibility in deepspeed #1427

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions diffsynth/models/ltx2_audio_vae.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,12 +55,14 @@ def waveform_to_mel(
waveform_sample_rate: int,
) -> torch.Tensor:
"""Convert waveform to log-mel spectrogram [batch, channels, time, n_mels]."""
original_dtype = waveform.dtype
waveform = self.resample_waveform(waveform, waveform_sample_rate, self.sample_rate)

mel = self.mel_transform(waveform)
self.mel_transform = self.mel_transform.to(dtype=torch.float32)

@gemini-code-assist gemini-code-assist Bot May 6, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Mutating the state of a module (self.mel_transform) during the forward pass is generally discouraged in PyTorch. This can lead to issues in distributed training environments like DeepSpeed (as mentioned in the PR title), especially with Stage 3 where parameters and buffers are managed and partitioned. Modifying them during forward can cause synchronization errors or unexpected behavior.

Additionally, the assignment self.mel_transform = self.mel_transform.to(...) is redundant because .to() on an nn.Module is an in-place operation. If you must ensure FP32 for FFT compatibility, consider ensuring the module is in the correct precision during initialization or setup, or at least avoid the redundant assignment.

Suggested change
self.mel_transform = self.mel_transform.to(dtype=torch.float32)
self.mel_transform.to(dtype=torch.float32)

mel = self.mel_transform(waveform.to(dtype=torch.float32))
mel = torch.log(torch.clamp(mel, min=1e-5))

mel = mel.to(device=waveform.device, dtype=waveform.dtype)
mel = mel.to(device=waveform.device, dtype=original_dtype)
return mel.permute(0, 1, 3, 2).contiguous()


Expand Down