Integrate RTX IO Support for NVIDIA open-gpu-kernel-modules for Enhanced LLM IO Efficiency

[ultranationalism]

As large models like DeepSeek are difficult to deploy on single consumer GPUs, many are opting for consumer GPUs paired with large memory (e.g., DDR5). This makes I/O efficiency critical, and integrating RTX IO into NVIDIA open-gpu-kernel-modules could significantly improve data transfer performance for such setups.like https://github.com/kvcache-ai/ktransformers