torch.compile and cuDNN errors block GPU inference
torch.compile and cuDNN errors block inference on NVIDIA GPUs, with compile=True and backend failures on Windows and Linux.
Tier II · Emerging Problems that are still building up.
3 public reports · first reported 3 June 2024 · last reported 1 August 2024
What's going wrong
Users running the ChatTTS text-to-speech project on NVIDIA GPUs report that inference fails once torch.compile is enabled with compile=True.
AI-assisted: written with the help of AI from the reports. The idea report has the full summary.
In the idea report
- The full summary.
- The products the 3 reports are about, and how many came each month.
- Related problems.
No business has been found with this problem yet, so the report has no contacts. All sales are final — no refunds. Pricing.