Hot Expert Reload on GPU is what this community needs

A huge favor to ask llama maintainers - please implement this feature. Even with one 3090 card there will be tangible improvements in decode speed on MOE models with moderate number of active parameters, like Qwen3.8-Flash-Next, Deepseek V4/V4.1 Flash, GLM 5.3 Flash. With 2x 3090 speeds will be quite close to the full offload of these models to VRAM. This will make these almost SOTA models really usable locally.