README.adoc

Run ComfyUI with ROCm on AMD GPU

English | 🀄 中文说明

View on <Docker Hub>

Tip	This image is using the latest ROCm 7, which is only recommended for Radeon RX 9000 series (RDNA 4) and Ryzen AI 300 series (RDNA 3.5) users.

Tip	If you are using RX 7000 series (RDNA 3) GPU, you may want to use the rocm6 image for a more stable experience.

Prequisites

• Make sure AMD GPU driver (kernel module amdgpu) is installed on your host system.

• Check if your GPU/CPU is supported by ROCm:

  • Supported GPU list

  • Supported Ryzen CPU (APU) list

  • Or follow the table below.

Essential Environment Variables

You need to add these configuration into the command of docker run or podman run below.

GPU Architecture	Configuration to add	GPU Model
RDNA 4 (gfx1201)	-e HSA_OVERRIDE_GFX_VERSION=12.0.1 \	Radeon AI PRO R9700
		Radeon RX 9070 XT
		Radeon RX 9070
		Radeon RX 9070 GRE
RDNA 4 (gfx1200)	-e HSA_OVERRIDE_GFX_VERSION=12.0.0 \	Radeon RX 9060 XT
		Radeon RX 9060
RDNA 3.5 (gfx1151)	-e HSA_OVERRIDE_GFX_VERSION=11.5.1 \ -e HIP_VISIBLE_DEVICES=0 \	Ryzen AI Max+ 395
		Ryzen AI Max 390
		Ryzen AI Max 385
RDNA 3.5 (gfx1150)	-e HSA_OVERRIDE_GFX_VERSION=11.5.0 \ -e HIP_VISIBLE_DEVICES=0 \	Ryzen AI 9 HX 375
		Ryzen AI 9 HX 370
		Ryzen AI 9 365
RDNA 3 (gfx1101)	-e HSA_OVERRIDE_GFX_VERSION=11.0.1 \	Radeon PRO W7700
		Radeon PRO V710
		Radeon RX 7800 XT
		Radeon RX 7700 XT
RDNA 3 (gfx1100)	-e HSA_OVERRIDE_GFX_VERSION=11.0.0 \	Radeon PRO W7900
		Radeon PRO W7800
		Radeon RX 7900 XTX
		Radeon RX 7900 XT
		Radeon RX 7900 GRE
RDNA 2 (gfx1030)	-e HSA_OVERRIDE_GFX_VERSION=10.3.0 \	Radeon PRO W6800
		Radeon PRO V620

• Radeon RX 6000 series may work, but there is no guarantee.

  • They are not officially supported by ROCm 7. But PRO GPUs of RDNA 2 (W6800, V620) are supported. They share the same build target gfx1030.

Reference: AMD GPU arch specs (Thanks to nhtua)

Run

Using Docker

mkdir -p \
#  storage \
  storage-models/models \
  storage-models/hf-hub \
  storage-models/torch-hub \
  storage-user/input \
  storage-user/output \
  storage-user/workflows

docker run \
  --name comfyui-rocm7 \
  --device=/dev/kfd --device=/dev/dri \
  --group-add=video --group-add=render \
  --ipc=host --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --security-opt label=disable \
  -p 8188:8188 \
#  -v "$(pwd)"/storage:/root \
  -v "$(pwd)"/storage-models/models:/root/ComfyUI/models \
  -v "$(pwd)"/storage-models/hf-hub:/root/.cache/huggingface/hub \
  -v "$(pwd)"/storage-models/torch-hub:/root/.cache/torch/hub \
  -v "$(pwd)"/storage-user/input:/root/ComfyUI/input \
  -v "$(pwd)"/storage-user/output:/root/ComfyUI/output \
  -v "$(pwd)"/storage-user/workflows:/root/ComfyUI/user/default/workflows \
  -e HSA_OVERRIDE_GFX_VERSION="" \
  -e CLI_ARGS="" \
  yanwk/comfyui-boot:rocm7

Using Podman

mkdir -p \
#  storage \
  storage-models/models \
  storage-models/hf-hub \
  storage-models/torch-hub \
  storage-user/input \
  storage-user/output \
  storage-user/workflows

podman run \
  --name comfyui-rocm7 \
  --device=/dev/kfd --device=/dev/dri \
  --group-add=video --group-add=render \
  --ipc=host --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --security-opt label=disable \
  -p 8188:8188 \
#  -v "$(pwd)"/storage:/root \
  -v "$(pwd)"/storage-models/models:/root/ComfyUI/models \
  -v "$(pwd)"/storage-models/hf-hub:/root/.cache/huggingface/hub \
  -v "$(pwd)"/storage-models/torch-hub:/root/.cache/torch/hub \
  -v "$(pwd)"/storage-user/input:/root/ComfyUI/input \
  -v "$(pwd)"/storage-user/output:/root/ComfyUI/output \
  -v "$(pwd)"/storage-user/workflows:/root/ComfyUI/user/default/workflows \
  -e HSA_OVERRIDE_GFX_VERSION="" \
  -e CLI_ARGS="" \
  docker.io/yanwk/comfyui-boot:rocm7

Once the app is loaded, visit http://localhost:8188/

Optional Environment Variables

You may also want to add more environment variables:

• Force ComfyUI to offload model weights from VRAM to RAM more frequently. Slows performance but reduce memory leaks (Source).

  • -e CLI_ARGS="--disable-smart-memory" \

• Disable internal memory fragment caching (to mitigate memory faults. Doc). (Thanks to SergeyFilippov)

  • -e HSA_DISABLE_FRAGMENT_ALLOCATOR=1 \

• Enable tunable operations (slower first run, faster subsequent runs. Doc1, Doc2). (Thanks to SergeyFilippov)

  • -e PYTORCH_TUNABLEOP_ENABLED=1 \