#1236 Will zero-copy be implemented on Strix Halo?

Beschreibung

Unfortunately, the current Strix Halo implementation mirrors the discrete GPU path: it fully loads expert weights into memory then creates additional expert cache slots, thereby exacerbating the already significant memory pressure. Since Strix Halo allows the GPU to directly access mapped host memory, employing the swap-in/swap-out strategies used for discrete GPUs is unnecessary and even detrimental. Are there plans to implement a proper zero-copy path tailored for unified memory in the future?