[Feature Request] NVLink bridging support for RTX20xx and RTX30xx cards

[mioiox]

Description:
Allow LTX to make use of two nVidia RTX GPUs linked via a NVLink bridge

Reasoning:
There are many old GPUs (RTX20 and RTX30 series) that can potentially be used for AI but are VRAM-limited - with 8, 10 or 12 GB. This is not really enough for most current models, and probably getting an AI model to work on that low amount of VRAM is a real challenge.

However, those cards do support an NVLink bridging of two GPUs for high-speed memory transfer. And even though current games do not really make use of that, it can be potentially leveraged by LTX. The current version of Blender (5.1) and its Cycles render engine do support NVLink for memory aggregation of two separate nVidia GPUs. I suppose the same concept can be reused for LTX.

Alternative approach:

• Replace the GPU
• Update the model to require less than 16 GB VRAM