Can you run it on mac?
Is mac supported for inference?
Ive tried running it using https://github.com/kijai/ComfyUI-HunyuanVideoWrapper/tree/main but kept running into mps issues. So as far as I'm aware not yet. However, https://github.com/BoBo0037/bobo-fun shows promise of supporting it on mac.
Is mac supported for inference?
yes technicality speaking, if you have over 90 gb of ram. but it will maybe take one day to generate a single video as the most tops an apple gpu has is 17 tops while an h100 has 204.9 TFLOPS
TLDR:
yes technicality speaking but very slow to run
I'd love to see MPS support on this model (https://developer.apple.com/documentation/metalperformanceshaders)
For example, M4 Max completes images at a 32-bit quant in under 4 minutes for black-forest-labs/FLUX.1-dev
I'd love to see MPS support on this model (https://developer.apple.com/documentation/metalperformanceshaders)
For example, M4 Max completes images at a 32-bit quant in under 4 minutes for black-forest-labs/FLUX.1-dev
A image model is a bit different than a video model
Successfully ran this on an 128GB M3 Max at work today.
This setup worked without issues (https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
~155s/it @ 100% GPU and ~100GB (V)RAM = 60-ish minutes for 73 frames (VAE decode took a few minutes, too).
Edit: 848x480
Successfully ran this on an 128GB M3 Max at work today.
This setup worked without issues (https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)~155s/it @ 100% GPU and ~100GB (V)RAM = 60-ish minutes for 73 frames (VAE decode took a few minutes, too).
Edit: 848x480
that's fast
FYI Tried it, but crashed after 54 minutes at 100% GPU and 128GB (V)RAM – it doesn't really seem faster than my comfyui setup.
2024-12-24 01:55:39.170 | INFO | hyvideo.diffusion.pipelines.pipeline_hunyuan_video:call:768 - Target latents shape: (1, 16, 13, 68, 120)
2024-12-24 02:49:26.550 | ERROR | hyvideo.diffusion.pipelines.pipeline_hunyuan_video:call:806 - Error in VAE decoding chunk 0: expected np.ndarray (got list)
/HunyuanVideo_MLX/hyvideo/vae/mlx_vae.py", line 74, in decode z = torch.from_numpy(z.astype(mx.float32).tolist())
that's fast
Is it? I only have LTXVideo for comparison atm – while quality seems lower (or harder to achieve?), it's about 10x faster.
Anyone any luck yet with https://github.com/kijai/ComfyUI-HunyuanVideoWrapper/tree/main? I also run into MPS issues in the HunyuanVideo Sampler node...