ComfyUI on MI60

Node-based Stable Diffusion UI with AMD MI60 GPU acceleration via ROCm.

Overview#

ComfyUI runs as a containerized service using a custom ROCm 5.7 image built for the MI60’s gfx906 architecture. It can run standalone on both GPUs or alongside vLLM in the image-chat configuration.

Quick Start#

# Build the container image
cd comfyui
make build

# Start ComfyUI (standalone, both GPUs)
make up

# Or use the image-chat config (vLLM on GPU0, ComfyUI on GPU1)
curl -X POST -H "Content-Type: application/json" \
  -d '{"config":"image-chat"}' http://localhost:9100/switch

Access the UI at http://localhost:8188.

Container Configuration#

The image-chat configuration runs ComfyUI on GPU1 alongside vLLM on GPU0:

comfyui:
  image: comfyui-rocm57:latest
  container_name: comfyui
  ports:
    - "8188:8188"
  devices:
    - /dev/kfd:/dev/kfd
    - /dev/dri/card2:/dev/dri/card2
    - /dev/dri/renderD129:/dev/dri/renderD129
  group_add:
    - "44"   # video
    - "992"  # render
  security_opt:
    - seccomp=unconfined
    - apparmor=unconfined
  cap_add:
    - SYS_PTRACE
  volumes:
    - /mnt/cache/comfyui/ComfyUI:/app/ComfyUI
  environment:
    - HSA_OVERRIDE_GFX_VERSION=9.0.6
    - HSA_ENABLE_SDMA=0
    - ROC_ENABLE_PRE_VEGA=1
    - HCC_AMDGPU_TARGET=gfx906
    - ROCR_VISIBLE_DEVICES=0
    - HIP_VISIBLE_DEVICES=0
  command: ["python3", "main.py", "--listen", "--fp8_e4m3fn-unet", "--fast"]

Key Environment Variables#

Variable	Value	Purpose
`HSA_OVERRIDE_GFX_VERSION`	9.0.6	Force gfx906 architecture detection
`HSA_ENABLE_SDMA`	0	Disable SDMA (stability fix)
`ROC_ENABLE_PRE_VEGA`	1	Enable pre-Vega compatibility
`HCC_AMDGPU_TARGET`	gfx906	Target architecture for HIP
`ROCR_VISIBLE_DEVICES`	0	Limit to single GPU in container

Command Line Options#

Option	Purpose
`--listen`	Accept connections from any host
`--fp8_e4m3fn-unet`	Enable FP8 for UNet (faster, less VRAM)
`--fast`	Enable fast mode optimizations

Storage Layout#

All data is stored on the host and mounted into containers:

/mnt/cache/comfyui/ComfyUI/
├── models/
│   ├── checkpoints/      # SD models (.safetensors)
│   ├── loras/            # LoRA adapters
│   ├── vae/              # VAE models
│   ├── controlnet/       # ControlNet models
│   └── upscale_models/   # Upscaler models
├── output/               # Generated images
├── input/                # Input images for img2img
└── custom_nodes/         # Installed extensions

Installing Models#

cd /mnt/cache/comfyui/ComfyUI/models/checkpoints

# SDXL base
wget https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors

# SD 1.5
wget https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned.safetensors

Installing Custom Nodes#

cd /mnt/cache/comfyui/ComfyUI/custom_nodes

# ComfyUI Manager (recommended)
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

# Restart ComfyUI to load new nodes

Troubleshooting#

“No HIP GPUs are available”#

Don’t rename devices - mount as-is (card2:/dev/dri/card2, not card2:/dev/dri/card0)
Mount both devices - need card AND renderD for each GPU
Set ROCR_VISIBLE_DEVICES=0 when container sees only one GPU