前言
随着 NVIDIA Blackwell 架构的问世，DGX Spark (Personal AI Supercomputer) 将桌面级 AI 算力推向了新的巅峰。这台设备搭载了 GB200/GB10 级别的 GPU 和 NVIDIA Grace CPU (ARM64)，并运行在最新的 CUDA 13 环境下。
然而，'最强硬件'往往伴随着'最难环境'。由于 Grace CPU 采用 ARM (aarch64) 架构，且 CUDA 13 过于前沿，传统的 PyTorch 安装方法极易失败。
本文将手把手教你如何在这台超级计算机上部署 Stable Diffusion 3.5 Large，并利用其 128GB 海量显存开启高性能模式，实现秒级、零延迟的创作体验。
硬件环境规格 (DGX Spark Specs)
本次部署基于 NVIDIA 最新一代个人 AI 超级计算机：
- Platform: NVIDIA DGX Spark (Personal AI Supercomputer)
- Architecture: NVIDIA Blackwell (GB10 / GB200 Tensor Core GPU)
- VRAM: 128GB HBM3e (Unified Memory Architecture)
- CPU: NVIDIA Grace CPU (144-core ARM64)
- Software Stack: NVIDIA AI Enterprise
- CUDA Version: CUDA 13.0
核心挑战与优势：挑战：标准 PyPI 源的 PyTorch 通常只适配 CUDA 11/12，直接安装会导致找不到 GPU。优势：128GB 显存允许我们将 SD3.5 的 20GB 模型 + 10GB T5 编码器永久驻留显存，消除模型加载时间。
第一步：构建兼容 CUDA 13 的基础环境
1.1 创建 Conda 环境
推荐使用 Python 3.11，它在 ARM 架构下的库支持最完善。
conda create -n sd-runtime python=3.11-y conda activate sd-runtime
1.2 安装适配 ARM + CUDA 13 的 PyTorch
这是最关键的一步。由于 CUDA 13 非常新，必须使用 NVIDIA 官方 PyPI 索引来获取正确的 wheel 包。
# 1. 确保 pip 是最新的
pip install --upgrade pip
# 2. 从 NVIDIA 官方源安装 PyTorch
# 注意：NVIDIA 的 index 通常会包含对最新 CUDA 版本的兼容包
pip install torch torchvision torchaudio --index-url https://pypi.nvidia.com
1.3 验证环境 (关键)
安装完成后，务必验证 PyTorch 是否成功链接到了 Blackwell GPU。
python -c "import torch; print(f'PyTorch Version: {torch.__version__}'); print(f'CUDA Available: {torch.cuda.is_available()}'); print(f'Device Name: {torch.cuda.get_device_name(0)}'); print(f'CUDA Version: {torch.version.cuda}')"
- 预期输出：
CUDA Available: TrueDevice Name: NVIDIA GB10 (或相关 Blackwell 代号)CUDA Version: 13.x
第二步：下载 Stable Diffusion 3.5 Large
SD3.5 是门控模型，请确保您已在 Hugging Face 签署协议并持有 Access Token。
# 1. 安装下载工具
pip install huggingface_hub
# 2. 配置环境变量 (使用国内镜像加速)
export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_ENABLE_HF_TRANSFER=1
export HF_TOKEN="你的_HF_Token_粘贴在这里"
# 3. 创建目录
mkdir -p /workspace/models/stabilityai/
# 4. 高速下载 (包含 Checkpoint 和 T5/CLIP 编码器)
huggingface-cli download stabilityai/stable-diffusion-3.5-large \
--token $HF_TOKEN \
--local-dir /workspace/models/stabilityai/stable-diffusion-3.5-large
第三步：部署 ComfyUI & 修复 ARM 依赖
3.1 克隆与安装
cd /workspace
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
# 安装常规依赖 (使用 --no-deps 防止破坏 Torch 环境)
pip install -r requirements.txt --no-deps
# 🔥 核心修复：补齐 ARM 架构下缺失的特殊库
# 如果缺少这些，ComfyUI 启动时会报错 "No module named kornia_rs" 或 "trampoline"
pip install trampoline kornia_rs mako comfyui-workflow-templates sqlalchemy alembic requests
3.2 挂载模型 (软链接方案)
为了利用 DGX Spark 的存储性能，我们使用软链接映射模型。
# 1. 链接主模型
ln -s /workspace/models/stabilityai/stable-diffusion-3.5-large/sd3.5_large.safetensors ./models/checkpoints/
# 2. 链接 CLIP 模型 (SD3 需要 3 个 CLIP)
# 注意：我们直接使用 FP16 精度的 T5，因为显存足够大
ln -s /workspace/models/stabilityai/stable-diffusion-3.5-large/text_encoders/clip_l.safetensors ./models/clip/
ln -s /workspace/models/stabilityai/stable-diffusion-3.5-large/text_encoders/clip_g.safetensors ./models/clip/
ln -s /workspace/models/stabilityai/stable-diffusion-3.5-large/text_encoders/t5xxl_fp16.safetensors ./models/clip/
第四步：高性能启动 (High VRAM Mode)
针对 128GB 显存，我们开启'疯狗模式'——让所有模型永久驻留显存，拒绝反复加载。
创建启动脚本 run_comfy.sh：
#!/bin/bash
WORKDIR="/workspace/ComfyUI"
PYTHON_PATH="/root/miniconda3/envs/sd-runtime/bin/python"
# 日志管理
LOG_DIR="$WORKDIR/logs"
mkdir -p "$LOG_DIR"
LOG_FILE="$LOG_DIR/server_$(date "+%Y-%m-%d_%H-%M-%S").log"
cd "$WORKDIR" || exit
echo "🚀 Starting ComfyUI on DGX Spark (Blackwell)..."
echo "⚡ Mode: High VRAM (Force Resident)"
# --highvram: 强制模型常驻显存
nohup "$PYTHON_PATH" main.py \
--listen 0.0.0.0 \
--port 8188 \
--highvram \
> "$LOG_FILE" 2>&1 &
echo "✅ Service PID: $!"
启动服务：chmod +x run_comfy.sh && ./run_comfy.sh
第五步：导入 SD3.5 标准工作流 (JSON)
SD3.5 采用了分离式文本编码器架构（CLIP-L + CLIP-G + T5），默认工作流无法正常出图。
请复制下方完整的 JSON 代码，保存为 sd3_workflow.json，直接拖入 ComfyUI 浏览器界面即可使用。此工作流已针对 1024x1024 分辨率和 DGX Spark 进行了优化。
{"last_node_id":10,"last_link_id":16,"nodes":[{"id":1,"type":"KSampler","pos":[960,190],"size":{"0":315,"1":262},"flags":{},"order":0,"mode":0,"inputs":[{"name":"model","type":"MODEL","link":1},{"name":"positive","type":"CONDITIONING","link":2},{"name":"negative","type":"CONDITIONING","link":3},{"name":"latent_image","type":"LATENT","link":4}],"outputs":[{"name":"LATENT","type":"LATENT","links":[5],"slot_index":0}],"properties":{"Node name for S&R":"KSampler"},"widgets_values":[1234567890,"randomize",30,4.5,"euler","sgm_uniform",1]},{"id":2,"type":"CheckpointLoaderSimple","pos":[50,50],"size":{"0":315,"1":98},"flags":{},"order":1,"mode":0,"outputs":[{"name":"MODEL","type":"MODEL","links":[1],"slot_index":0},{"name":"CLIP","type":"CLIP","links":[],"slot_index":1},{"name":"VAE","type":"VAE","links":[6],"slot_index":2}],"properties":{"Node name for S&R":"CheckpointLoaderSimple"},"widgets_values":["sd3.5_large.safetensors"]},{"id":3,"type":"TripleCLIPLoader","pos":[50,200],"size":{"0":315,"1":150},"flags":{},"order":2,"mode":0,"outputs":[{"name":"CLIP","type":"CLIP","links":[7,8],"slot_index":0}],"properties":{"Node name for S&R":"TripleCLIPLoader"},"widgets_values":["clip_l.safetensors","clip_g.safetensors","t5xxl_fp16.safetensors"]},{"id":4,"type":"CLIPTextEncode","pos":[450,100],"size":{"0":425,"1":180},"flags":{},"order":3,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":7}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","links":[2],"slot_index":0}],"properties":{"Node name for S&R":"CLIPTextEncode"},"widgets_values":["a highly detailed cyberpunk city, neon lights, 8k resolution, cinematic lighting, masterpiece, sharp focus, ray tracing, nvidia blackwell style"],"color":"#233","bgcolor":"#353"},{"id":5,"type":"CLIPTextEncode","pos":[450,350],"size":{"0":425,"1":180},"flags":{},"order":4,"mode":0,"inputs":[{"name":"clip","type":"CLIP","link":8}],"outputs":[{"name":"CONDITIONING","type":"CONDITIONING","links":[3],"slot_index":0}],"properties":{"Node name for S&R":"CLIPTextEncode"},"widgets_values":["low quality, blurry, ugly, bad anatomy, watermark, text, jpeg artifacts"],"color":"#332","bgcolor":"#533"},{"id":6,"type":"EmptyLatentImage","pos":[50,450],"size":{"0":315,"1":106},"flags":{},"order":5,"mode":0,"outputs":[{"name":"LATENT","type":"LATENT","links":[4],"slot_index":0}],"properties":{"Node name for S&R":"EmptyLatentImage"},"widgets_values":[1024,1024,1]},{"id":7,"type":"VAEDecode","pos":[1300,190],"size":{"0":210,"1":46},"flags":{},"order":6,"mode":0,"inputs":[{"name":"samples","type":"LATENT","link":5},{"name":"vae","type":"VAE","link":6}],"outputs":[{"name":"IMAGE","type":"IMAGE","links":[9],"slot_index":0}],"properties":{"Node name for S&R":"VAEDecode"}},{"id":8,"type":"SaveImage","pos":[1550,190],"size":{"0":500,"1":600},"flags":{},"order":7,"mode":0,"inputs":[{"name":"images","type":"IMAGE","link":9}],"properties":{"Node name for S&R":"SaveImage"},"widgets_values":["ComfyUI"]}],"links":[[1,2,0,1,0,"MODEL"],[2,4,0,1,1,"CONDITIONING"],[3,5,0,1,2,"CONDITIONING"],[4,6,0,1,3,"LATENT"],[5,1,0,7,0,"LATENT"],[6,2,2,7,1,"VAE"],[7,3,0,4,0,"CLIP"],[8,3,0,5,0,"CLIP"],[9,7,0,8,0,"IMAGE"]],"groups":[],"config":{},"extra":{},"version":0.4}
工作流配置说明：
- TripleCLIPLoader: 这是核心。我们明确指定了
clip_l,clip_g和t5xxl_fp16。如果这三个没选对，生成的图片将无法理解你的 Prompt。 - Scheduler: sgm_uniform: Stable Diffusion 3 必须使用 SGM Uniform 调度器，否则画面会充满噪点。
- Resolution: 默认设为 1024x1024。在 Blackwell 上，你可以尝试将其改为 2048x2048 进行原生 4K 生成。
性能总结
在 DGX Spark 上完成上述部署后，性能表现如下：
- 模型加载: 首次启动后，模型将永久驻留显存 (High VRAM Mode)。
- 生成速度: 1024x1024 分辨率下，实现秒级出图。
- 并发能力: 得益于 128GB 显存，你可以同时开启
Batch Size = 4甚至更高，或者在后台同时运行一个 70B 的大语言模型而不发生 OOM。
