Support for older NVIDIA GPUs (Pascal architecture - GTX 1080, etc.) with PyTorch 1.13

[HenishAhmed]

Description

The current version requires PyTorch 2.x (CUDA 11.8+), which dropped support for older NVIDIA GPU architectures. Specifically, Pascal architecture GPUs (like GTX 1080) are no longer supported as they only have Compute Capability 6.1, while PyTorch 2.0+ requires minimum Compute Capability 7.0+.

These are still capable cards for TTS inference, and even thought their a bit old it's better than using the CPU.

Proposed Solution

Add support for PyTorch 1.13.x as an optional fallback configuration for users with older GPUs. This was the last PyTorch version to support Compute Capability 6.1.

Suggested implementation:

• Add a cu117 or legacy_cu configuration option in lib/conf.py (line 66-98 torch_matrix)
• Set PyTorch 1.13.1 for older CUDA installations
• Document compatibility and any potential feature limitations

Current Behavior

• Error: CUDA out of memory or GPU detection failure with Pascal GPUs
• Fallback to CPU-only mode (very slow)

Expected Behavior

• Allow users with Pascal GPUs to use CUDA acceleration via PyTorch 1.13
• Provide documentation on which TTS engines are compatible with 1.13

Additional Context

• PyTorch 1.13 still supports most modern TTS engines (Coqui, XTTS, etc.)
• Users with older hardware would appreciate this option rather than complete exclusion
• Many users still have functional GTX 1080s and similar cards (I am one of them)

I really hope you could help me, I really appreciate your hard work 😊

[ROBERT-MCDOWELL]

we never supported torch 1.x since the tts engine we are using requires min torch 2.3.x. not only that, many components require torch, it's extremely complex (already) to keep the version of each component compatible. for now we are managing only coqui-tts but planned to add more like piper-tts etc.. which render the requirements more complex. this said,
you are free to find fixed version of each component in requirements.txt, starting with coqui-tts max version to 0.17.x

[HenishAhmed]

Update (4 days later): I tried what you suggested and successfully ran CUDA
on my GTX 1080. It all went smoothly and I had no problems running the program.
Thanks again for the guidance! 😊