Comment by storus

Strix Halo can only allocate 96GB RAM to the GPU. So GPT-OSS 120B can be ran only at Q6 at best (but activations would need to be partially stored in the CPU mem then).