make 'Vae Patched' the default mode for safer decoding

Updated the INPUT_TYPES definition to set "Vae Patched" as the default selection.

Why: The "Original" VAE mode often causes black images or NaN errors in mixed-precision workflows (like Flux, Z_Image, or FP16 contexts) on certain hardware.

Fix: By defaulting to "Vae Patched," the node now automatically handles precision casting (forcing Float32 when necessary) to ensure successful image decoding out of the box, while still allowing advanced users to switch back to "Original" if needed.

Add legacy GPU support (SM < 8.0) and fix dynamic device switching

This update introduces a robust "Runtime Device Guard" and intelligent hardware detection to solve crashes on older GPUs and multi-GPU workflows.

Key Changes:

Legacy GPU Support (SM < 8.0): Added automatic detection for older architectures (Pascal/Volta/Turing). The node now forcefully disables Flash Attention and xFormers on these cards to prevent crashes, falling back to standard PyTorch attention with safe Float32 precision.

Dynamic Device Switching Fix: Implemented a "Self-Healing" Runtime Guard that intercepts the model's forward() method. It detects where the input tensors are coming from and instantly moves the model weights to match the target device. This fixes RuntimeError: Expected all tensors to be on the same device when switching between cuda:0 and cuda:1 across multiple runs or samplers.

Precision Safety: Added logic to preserve the original output dtype (Half/Float16) after safe Float32 calculations, resolving float != c10::Half errors in subsequent model layers.

Universal Input Caster: Updated the input scanner to check both *args and **kwargs for tensors, fixing AttributeError: 'NoneType' crashes on custom architectures that pass hidden states via keywords.

Making the default selections

making the default selections set to help beginners in ComfyUI

First Release

This custom node gives you total control over where your models run and where they live in memory. It solves common (OOM) errors, enables multi-GPU workflows, and fixes persistent crashes when trying to run modern workflows (like Z_Image/Flux) on CPUs or secondary GPUs.