AI Recipes для GPU серверов MetalNode

cd /work
mkdir -p ai-lab
cd /work/ai-lab

apt update
apt install -y python3-venv python3-pip git wget

python3 -m venv .venv-cu128
source .venv-cu128/bin/activate

python -m pip install -U pip setuptools wheel
pip install --no-cache-dir --retries 20 --timeout 300 torch torchvision --index-url https://download.pytorch.org/whl/cu128
pip install --no-cache-dir --retries 1 --timeout 10 torchaudio --index-url https://download.pytorch.org/whl/cu128 || true
pip install diffusers transformers accelerate safetensors sentencepiece protobuf pillow

python - <<'PY'
import torch
print("torch:", torch.__version__)
print("cuda available:", torch.cuda.is_available())
print("gpu count:", torch.cuda.device_count())
if torch.cuda.is_available():
    print("gpu name:", torch.cuda.get_device_name(0))
    x = torch.randn(64, 64, device='cuda')
    y = torch.matmul(x, x)
    print("matmul ok:", y.shape, y.dtype, y.device)
PY

cat > /work/ai-lab/gen_image.py <<'PY'
import time
import torch
from diffusers import AutoPipelineForText2Image

MODEL_ID = "stabilityai/sdxl-turbo"
OUT_PATH = "/work/ai-lab/out_sdxl_turbo.png"

prompt = "a futuristic bare metal GPU server room, cinematic lighting, ultra detailed, realistic, blue neon accents"

print("Loading model:", MODEL_ID)
t0 = time.time()

pipe = AutoPipelineForText2Image.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe = pipe.to("cuda")

print("Model loaded in %.1f sec" % (time.time() - t0))

t1 = time.time()
image = pipe(
    prompt=prompt,
    num_inference_steps=4,
    guidance_scale=0.0
).images[0]

image.save(OUT_PATH)

print("Saved:", OUT_PATH)
print("Generation time: %.1f sec" % (time.time() - t1))
PY

time python /work/ai-lab/gen_image.py
ls -lh /work/ai-lab/out_sdxl_turbo.png

set -e

echo "===== MetalNode AI Studio Base install ====="

apt update
apt install -y git python3-venv python3-pip ffmpeg curl tmux

mkdir -p /work/ai
cd /work/ai

python3 -m venv venv
source venv/bin/activate

pip install --upgrade pip

echo "Installing PyTorch CUDA 12.8..."
pip install --no-cache-dir --retries 20 --timeout 300 torch torchvision \
  --index-url https://download.pytorch.org/whl/cu128

echo "Installing torchaudio (optional)..."
pip install --no-cache-dir --retries 1 --timeout 10 torchaudio \
  --index-url https://download.pytorch.org/whl/cu128 || true

echo "Installing xFormers CUDA 12.8..."
pip install xformers \
  --index-url https://download.pytorch.org/whl/cu128

echo "Installing AI libraries..."
pip install \
  diffusers \
  transformers \
  accelerate \
  safetensors \
  huggingface_hub \
  opencv-python \
  insightface \
  onnxruntime \
  piexif \
  segment-anything

echo "Cloning ComfyUI..."
cd /work
if [ ! -d /work/ComfyUI ]; then
  git clone https://github.com/comfyanonymous/ComfyUI.git
fi

cd /work/ComfyUI
source /work/ai/venv/bin/activate

pip install -r requirements.txt

echo "Installing ComfyUI Manager..."
mkdir -p custom_nodes
cd custom_nodes

[ -d ComfyUI-Manager ] || git clone https://github.com/ltdrdata/ComfyUI-Manager.git

echo "Installing useful AI nodes..."
[ -d ComfyUI-AnimateDiff-Evolved ] || git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved.git
[ -d ComfyUI-Impact-Pack ] || git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack.git
[ -d ComfyUI_IPAdapter_plus ] || git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus.git
[ -d ComfyUI-ReActor ] || git clone https://github.com/Gourieff/ComfyUI-ReActor.git
[ -d ComfyUI_UltimateSDUpscale ] || git clone https://github.com/ssitu/ComfyUI_UltimateSDUpscale.git

cd /work/ComfyUI

echo "Creating model folders..."
mkdir -p models/checkpoints
mkdir -p models/loras
mkdir -p models/controlnet
mkdir -p models/upscale_models
mkdir -p models/clip
mkdir -p models/vae
mkdir -p models/animatediff_models

echo "Downloading SDXL Turbo checkpoint..."
source /work/ai/venv/bin/activate
if [ ! -f models/checkpoints/sdxl-turbo.safetensors ]; then
  hf download stabilityai/sdxl-turbo \
    sd_xl_turbo_1.0_fp16.safetensors \
    --local-dir models/checkpoints
  mv -f models/checkpoints/sd_xl_turbo_1.0_fp16.safetensors \
    models/checkpoints/sdxl-turbo.safetensors
fi

echo "Downloading SDXL-Lightning 4-step LoRA..."
mkdir -p models/loras/SDXL-Lightning
if [ ! -f models/loras/SDXL-Lightning/sdxl_lightning_4step_lora.safetensors ]; then
  hf download ByteDance/SDXL-Lightning \
    sdxl_lightning_4step_lora.safetensors \
    --local-dir models/loras/SDXL-Lightning
fi

echo "Creating helper scripts..."
mkdir -p /work/bin

cat > /work/bin/start-comfy.sh <<'SH'
#!/usr/bin/env bash
set -e
cd /work/ComfyUI
source /work/ai/venv/bin/activate
exec python main.py --listen --port 8188
SH
chmod +x /work/bin/start-comfy.sh

cat > /work/bin/run-comfy-tmux.sh <<'SH'
#!/usr/bin/env bash
set -e
SESSION="comfy"
if tmux has-session -t "$SESSION" 2>/dev/null; then
  echo "tmux session '$SESSION' already exists"
  exit 0
fi
tmux new-session -d -s "$SESSION" "/work/bin/start-comfy.sh"
echo "Started ComfyUI in tmux session '$SESSION'"
echo "Attach with: /work/bin/attach-comfy.sh"
SH
chmod +x /work/bin/run-comfy-tmux.sh

cat > /work/bin/attach-comfy.sh <<'SH'
#!/usr/bin/env bash
exec tmux attach -t comfy
SH
chmod +x /work/bin/attach-comfy.sh

cat > /work/bin/stop-comfy.sh <<'SH'
#!/usr/bin/env bash
tmux kill-session -t comfy 2>/dev/null || true
echo "Stopped tmux session 'comfy'"
SH
chmod +x /work/bin/stop-comfy.sh

cat > /work/bin/check-ai-stack.sh <<'SH'
#!/usr/bin/env bash
set -e
source /work/ai/venv/bin/activate
python - <<'PY'
import torch
print("torch:", torch.__version__)
print("cuda available:", torch.cuda.is_available())
print("gpu count:", torch.cuda.device_count())
if torch.cuda.is_available():
    print("gpu name:", torch.cuda.get_device_name(0))
    x = torch.randn(64, 64, device='cuda')
    y = torch.matmul(x, x)
    print("matmul ok:", y.shape, y.dtype, y.device)
PY

echo
echo "Models:"
ls -lh /work/ComfyUI/models/checkpoints || true
echo
ls -lh /work/ComfyUI/models/loras/SDXL-Lightning || true
SH
chmod +x /work/bin/check-ai-stack.sh

echo ""
echo "===== Install complete ====="
echo ""
echo "Start ComfyUI in background:"
echo "/work/bin/run-comfy-tmux.sh"
echo ""
echo "Attach to running session:"
echo "/work/bin/attach-comfy.sh"
echo ""
echo "Stop ComfyUI:"
echo "/work/bin/stop-comfy.sh"
echo ""
echo "Check AI stack:"
echo "/work/bin/check-ai-stack.sh"
echo ""
echo "Open via SSH tunnel:"
echo "ssh -L 8188:localhost:8188 root@SERVER -p PORT"
echo "http://localhost:8188"
echo ""
echo "Installed models:"
echo " - SDXL Turbo checkpoint"
echo " - SDXL-Lightning 4-step LoRA"
echo ""
echo "Notes:"
echo " - AnimateDiff node is installed, but motion models are NOT downloaded by default"
echo " - ReActor and Impact Pack should load because cv2/insightface/onnxruntime/piexif/segment-anything are installed"