Mac M-series (M3 MX 48GB) with unified memory should be the ideal hardware for large video models — no VRAM bottleneck, massive memory pool. But MPS lacks torch.compile support, IC-LoRAs are CUDA-only, and every 14B model pushes GPU temps to 86°C+ with no viable path to production output. LTX 2.3 and Wan 2.2 both need Mac-native Metal optimization. The hardware is capable — the software stack isn't. Apple Silicon users represent a significant and growing user base. Please prioritize MPS/Metal parity.
Mac M-series (M3 MX 48GB) with unified memory should be the ideal hardware for large video models — no VRAM bottleneck, massive memory pool. But MPS lacks torch.compile support, IC-LoRAs are CUDA-only, and every 14B model pushes GPU temps to 86°C+ with no viable path to production output. LTX 2.3 and Wan 2.2 both need Mac-native Metal optimization. The hardware is capable — the software stack isn't. Apple Silicon users represent a significant and growing user base. Please prioritize MPS/Metal parity.