Adaptive OS-based support for RTX series

[vecnode]

• This needs to be implemented, making it faster
• also check diffusers

[vecnode]

Quick status + next steps for this RTX support work:

What is in place already

• Core diffusion logic is being refactored toward a more modular pipeline, which makes it easier to plug in device-specific execution strategies without touching the math itself.
• Configuration is being centralized, so device selection and RTX-related options can live in one place instead of being threaded ad-hoc through the code.

What remains to fully resolve this issue

• Design a small, OS-aware device abstraction layer that:
  • Detects available RTX-capable GPUs across Linux/macOS/Windows where applicable.
  • Encapsulates CUDA / driver / library checks and error messages, without leaking into diffusion math code.
• Define per-OS defaults for things like memory growth, precision modes, and backend flags that are safe for RTX hardware.
• Wire this abstraction into the run configuration so that users can either:
  • Let the system auto-select an RTX device when available, or
  • Explicitly choose devices via config/CLI, with consistent behavior across OSes.
• Add minimal docs on supported OS + RTX combinations and known caveats.

The goal is to keep all RTX and OS-specific concerns inside this device layer, so the diffusion math remains clean and portable while still taking advantage of RTX when present.