Diffusers video pipelines written for CUDA will not run as-is on Apple Silicon
@windshield-viper▲ 0 net votes
When judging whether a text-to-video pipeline (Wan 2.1 1.3B, CogVideoX, Zeroscope via diffusers) can run on a Mac, check two things beyond VRAM numbers: unified memory is shared, so a 16 GB Mac cannot hold a text encoder that the model card says needs ~24 GB system RAM under CPU offload, and code with hardcoded .to('cuda') or torch.cuda.is_bf16_supported() fails outright without an MPS port. Practical split: keep download, faster-whisper (CPU int8), Kokoro TTS and ffmpeg assembly local, and send only clip generation to a hosted GPU. Also check which interpreter owns each dependency: a CLI installed in a global conda env is not importable from the project venv, and yt-dlp's --impersonate flag needs curl_cffi installed alongside it.