MiniMax H3 FL2VA successfully running on M5 MacBook Air with 32GB unified memory

[Atsushi888]

Hi, I wanted to share a successful MiniMax H3 test on a MacBook Air.
Environment

• MacBook Air 15-inch
• Apple M5
• 32GB unified memory
• VPIPE on macOS
• MiniMax H3 FL2VA
• 8-bit quantized model
• i8_gemm: true
  Successful generation
• Resolution: 960 × 544
• DiT steps: 8
• GPU utilization during generation: approximately 99%
• Generation completed successfully in approximately 21–24 minutes in my tests
• Video and audio were successfully generated and saved as MP4
• The Web UI pipeline also completed successfully
  Getting here took some trial and error.
  I initially prepared a different quantization/model configuration and eventually had to prepare the model again as FL2VA 8-bit. I also ran into model-registry/path issues in the Web UI. The model picker did not recognize my prepared FL2VA model correctly, but entering the absolute path to the existing FL2VA-8bit model in model-select allowed the pipeline to run successfully.
  Another thing that initially confused me was the Web UI output location. The generated file was eventually found under the VPIPE work directory's sandbox directory.
  Despite those setup difficulties, once everything was configured correctly, MiniMax H3 ran all the way through on the 32GB MacBook Air and produced a complete video with audio.
  So I can confirm that MiniMax H3 FL2VA is usable on an M5 MacBook Air with 32GB unified memory.
  Thank you for making this possible on Apple Silicon. I hope these results are useful to other MacBook Air users.
  My next tests will probably be Turbo LoRA and image-to-video.

[slicedPear]

I've discovered the same path issues in the pipeline configs when using an external drive. The prepare-minimax-h3-8bit pipeline downloaded the HF model @ nearly 200Gb but failed to quantise and just stopped when I re-ran. Just went into the node config and put the absolute path and it's running now, so let's see.

I'll keep an eye out for path issues in the generation pipeline, thanks.