Running ComfyUI on AMD GPUs with ROCm (Pop!_OS / Ubuntu)

wget https://repo.radeon.com/rocm/rocm.gpg.key -qO - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null

echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/latest $(lsb_release -cs 2>/dev/null) main" | sudo tee --append /etc/apt/sources.list.d/rocm.list

echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600

sudo apt update
sudo apt install rocm

sudo usermod -a -G render $USER

rocminfo

comfy(){
  ai && \
  HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  HSA_ENABLE_SDMA=0 \
  PYTORCH_HIP_ALLOC_CONF="garbage_collection_threshold:0.8,max_split_size_mb:512" \
  python /home/$USER/Documents/ComfyUI/main.py \
  --disable-pinned-memory
}

torchcheck() {
  source ~/.virtualenvs/ai/bin/activate
  python - <<'PY'
import torch
print("torch:", torch.__version__)
print("cuda available:", torch.cuda.is_available())
print("hip:", getattr(torch.version, "hip", None))
if torch.cuda.is_available():
    print("device:", torch.cuda.get_device_name(0))
PY
}

detect_rocm() {
  if [[ -f /opt/rocm/.info/version ]]; then
    cat /opt/rocm/.info/version
  else
    rocminfo | head
  fi
}

resetcomfy_rocm72() (
  LOGFILE="$HOME/resetcomfy_rocm72.log"

  {
    echo "================================================="
    echo " resetcomfy_rocm72 started at $(date)"
    echo " Log file: $LOGFILE"
    echo "================================================="

    set -x

    echo "⚠️  This will DELETE ComfyUI and the ai virtualenv. Ctrl+C to cancel."
    sleep 3

    # --- Cleanup ---
    rm -rf /home/$USER/Documents/ComfyUI
    rm -rf /home/$USER/.virtualenvs/ai
    rm -rf /home/$USER/config/comfy

    # --- Virtualenv ---
    python3 -m venv ~/.virtualenvs/ai
    source ~/.virtualenvs/ai/bin/activate

    pip install -U pip wheel

    # --- Clone ComfyUI ---
    cd /home/$USER/Documents
    git clone https://github.com/comfyanonymous/ComfyUI.git
    cd ComfyUI

    # --- Install ComfyUI requirements WITHOUT deps (never pull CUDA torch) ---
    pip install --no-deps -r requirements.txt

    # =========================================================
    # FORCE ROCm TORCH 7.2 (KNOWN GOOD)
    # =========================================================

    # Remove CUDA / NVIDIA torch stack
    pip uninstall -y torch torchvision torchaudio triton cuda-pathfinder >/dev/null 2>&1 || true

    # Remove any installed nvidia-*-cu12 wheels safely
    NVID_PKGS="$(pip freeze | sed -n 's/^\(nvidia-.*-cu12\)==.*/\1/p')"
    if [[ -n "$NVID_PKGS" ]]; then
      echo "$NVID_PKGS" | xargs -r pip uninstall -y
    fi

    # Install ROCm 7.2 wheels (Python 3.12)
    cd /tmp
    rm -rf rocm72_torch && mkdir rocm72_torch && cd rocm72_torch

    curl -L -O "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torch-2.9.1%2Brocm7.2.0.lw.git7e1940d4-cp312-cp312-linux_x86_64.whl"
    curl -L -O "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchvision-0.24.0%2Brocm7.2.0.gitb919bd0c-cp312-cp312-linux_x86_64.whl"
    curl -L -O "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchaudio-2.9.0%2Brocm7.2.0.gite3c6ee2b-cp312-cp312-linux_x86_64.whl"

    pip install --no-deps ./torch-*.whl ./torchvision-*.whl ./torchaudio-*.whl

    # --- Torch sanity check ---
    python - <<'PY'
import torch
print("torch:", torch.__version__)
print("hip:", getattr(torch.version, "hip", None))
PY

    # =========================================================
    # COMFYUI-MANAGER PROTECTION (STOP CUDA REINSTALL)
    # =========================================================

    mkdir -p /home/$USER/Documents/ComfyUI/user/__manager

    cat > /home/$USER/Documents/ComfyUI/user/__manager/pip_blacklist.list <<'EOF'
torch
torchvision
torchaudio
triton
cuda-pathfinder
nvidia-cublas-cu12
nvidia-cuda-cupti-cu12
nvidia-cuda-nvrtc-cu12
nvidia-cuda-runtime-cu12
nvidia-cudnn-cu12
nvidia-cufft-cu12
nvidia-cufile-cu12
nvidia-curand-cu12
nvidia-cusolver-cu12
nvidia-cusparse-cu12
nvidia-cusparselt-cu12
nvidia-nccl-cu12
nvidia-nvjitlink-cu12
nvidia-nvshmem-cu12
nvidia-nvtx-cu12
EOF

    # --- Install ComfyUI Manager ---
    cd /home/$USER/Documents/ComfyUI/custom_nodes
    git clone https://github.com/ltdrdata/ComfyUI-Manager comfyui-manager
    cd comfyui-manager
    pip install --no-deps -r requirements.txt

    # =========================================================
    # PURE-PYTHON DEPS (NO CUDA, NO TORCH)
    # =========================================================

    pip install -U \
      filelock sympy networkx jinja2 fsspec setuptools typing-extensions \
      aiohttp yarl multidict frozenlist aiosignal attrs \
      alembic SQLAlchemy greenlet Mako \
      pydantic pydantic-settings annotated-types pydantic-core typing-inspection \
      certifi charset-normalizer urllib3 \
      httpx httpcore regex trampoline \
      kornia_rs \
      comfyui-workflow-templates \
      comfyui-frontend-package

    echo "================================================="
    echo " ✅ SUCCESS at $(date)"
    echo "================================================="
  } 2>&1 | tee -a "$LOGFILE"
)

hfd() {
    if [ -z "$1" ]; then
        echo "Usage: hfd <repo_id> [optional_include_pattern]"
        return 1
    fi

    local VENV_PATH="/home/$USER/.virtualenvs/home/bin/hf"

    if [ -z "$2" ]; then
        # If no second argument, download the whole repo
        $VENV_PATH download "$1" --local-dir "./$(basename $1)"
    else
        # If second argument provided, download specific file/pattern
        $VENV_PATH download "$1" --include "$2" --local-dir "./$(basename $1)"
    fi
}