Minimax H3 - AMD HIP and MULTIGPU tips

It works very well for me and has been stable and error-free. Follow the installation instructions on GitHub.

After installing, make sure to update everything to the latest version using the included batch files:

comfyui-rocm-updater.bat rocm-pytorch-package-updater.bat

Then edit:

comfyui-rocm.bat

Find the line:

set PARAMS=

and add:

--use-sage-attention

at the end.

2. Download the Q8_0 MiniMax H3 models

I recommend using the Q8_0 models.

The INT8 versions also work and are roughly as fast as Q8, but in my testing they require significantly more VRAM with no noticeable quality benefit .

I used the Unsloth versions:

https://huggingface.co/unsloth/MiniMax-H3-GGUF/tree/main

Specifically:

minimax_h3_fl2va_pruned-Q8_0.gguf minimax_h3_ref2va_pruned-Q8_0.gguf

If you only have one 24 GB GPU , I suggest starting with Q5_0 quantization instead.

3. Use the recommended CLIP in NVFP4

Download the recommended CLIP:

qwen3vl_32b_minimax_h3_nvfp4_awq

This was the main point of confusion for me.

I'm familiar with NVFP4 from the LLM world, and because NVFP4 is associated with NVIDIA hardware, I initially assumed it wouldn't work on AMD.

I tried several GGUF CLIP models and managed to get them working only for T2V . For the other workflows, I had no luck.

The important part is:

You do NOT need a GGUF CLIP here. NVFP4 works on AMD in this setup.

The included comfy-kitchen module handles the quantized model transparently, so the recommended NVFP4 CLIP can be used directly.

4. Install the required custom nodes

Using ComfyUI Manager , install:

comfyui-multigpu ComfyUI-GGUF

After installing comfyui-multigpu , you'll get several new loader nodes.

For the GGUF model, use:

UnetLoaderGGUFDisTorch2MultiGpu

On my 2× RX 7900 XTX 24 GB setup , I use:

virtual_vram_gb = 20

This value is specific to my hardware configuration and should not be treated as a universal recommendation. If you're using GPUs with a different amount of VRAM, you'll likely need to adjust it accordingly.

In my setup, this lets the second 7900 XTX act as a very fast cache/offload device for the primary 7900 XTX doing the main computation.

5. Optional: Turbo LoRAs

You can add LoRAs on top if you want.

The MiniMax H3 LoRA ecosystem is evolving quickly, but these were among the first Turbo LoRAs available:

https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main

They allow you to generate with around 4 or 8 steps instead of the default 20 , depending on the LoRA and workflow.

This setup is currently working very well for me on 2× RX 7900 XTX 24 GB , especially with the multi-GPU loader.

I'll attach the workflow I used as well.

Checkpoint · Hugging Face · lightx2v/Minimax-h3-Turbo · Minimax H3 Turbo