Kijai

/

MiniMax-H3-experimental

like 201

it doesn't mater, the vae loader can't do quantized vae:s yet and by default for minimax uses fp16, it is by design, even if u use fp8 the speed gain is minimal, 1s maybe,
compute dtype is set in comfy/sd.py : https://github.com/Comfy-Org/ComfyUI/blob/2340099d93305bfdf4eaa29e9f8d32ec92d3035f/comfy/sd.py#L942-L949

elif "decoder.transformer_blocks.0.scale1" in sd and "encoder.down.5.block.0.conv1.weight" in sd:  # MiniMax H3 video VAE
    self.first_stage_model = comfy.ldm.minimax.vae.MiniMaxH3VideoVAE(num_layers=n_layers)
    ...
    self.working_dtypes = [torch.float16, torch.float32]   

fp8 VAE is not a good idea, but the int8_convrot VAE works pretty well and is 1.5x faster (requires comfyui to be updated to use)

@rzgar it wasn't real mixed fp8 anyway, just mixed static fp8 with bf16, was one of my experiments, outdated now