Running LingBot-Video on a 64GB Apple Silicon Mac
LingBot-Video is an Apache 2.0 text-to-video model that can run on your own machine instead of charging by the second through a . Its official software setup, however, assumes an NVIDIA and CUDA throughout. The 30-billion-parameter MoE version is too large for this computer: its main model alone needs about 60GB in bf16, while a 64GB Mac makes only about 55.7GB available to graphics work.
The smaller 1.3-billion-parameter version is practical and has roughly 12GB of weights on disk. Running it on required replacing the CUDA-only PyTorch test build with standard PyTorch 2.13 and MPS, then using SDPA instead of -3. Because decord has no arm64 Mac package, eva-decord served as a compatible replacement.
A further problem was that MPS could stop the program when a bf16 Linear layer received fp32 input, rather than showing a normal error, so the data types had to be matched.
Key points
- The 30-billion-parameter model exceeds the limit of a 64GB Mac.
- The smaller 1.3-billion-parameter model has about 12GB of weights.
- SDPA replaced the NVIDIA-only -3 component.
- eva-decord replaced decord because the latter lacks an arm64 Mac package.
- The bf16 and fp32 data types must be matched to prevent an MPS crash.