Fix dtype mismatch in AudioVAE.encode for bf16/fp16 audio VAE

This commit is contained in:
Refuelko 2026-07-06 23:08:30 -07:00 committed by GitHub
parent 7cf4e78335
commit ad3797ad4f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -154,6 +154,10 @@ class AudioVAE(torch.nn.Module):
waveform, waveform_sample_rate, device=waveform.device
)
# Cast mel spectrogram to match encoder weight dtype (bf16 VAE + float32 audio input)
encoder_dtype = next(self.autoencoder.encoder.parameters()).dtype
mel_spec = mel_spec.to(dtype=encoder_dtype)
latents = self.autoencoder.encode(mel_spec)
posterior = DiagonalGaussianDistribution(latents)
latent_mode = posterior.mode()