A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 14:13 UTC
Minor cleanup for Voxtral (#34247)
Signed-off-by: Andy Lo <andy@mistral.ai>
Andy Lo committed
Feb 10, 2026 at 18:18 UTC
ae871ca9234be3f6cb6966d998e51a7cb672f912
1 file changed
+3
-1
vllm/model_executor/models/voxtral.py
+3
-1
@@ -792,7 +792,9 @@ class VoxtralEncoderModel(nn.Module):
792
audio_waveforms: torch.Tensor,
793
) -> torch.Tensor:
794
input_dtype = audio_waveforms.dtype
795
- window = torch.hann_window(self.config.window_size).to(audio_waveforms.device)
795
+ window = torch.hann_window(
796
+ self.config.window_size, device=audio_waveforms.device
797
+ )
798
stft = torch.stft(
799
audio_waveforms,
800
self.config.window_size,