Stable Diffusion 宿主机部署痛点拆解 + Docker 三方案（Run/Compose/Dockerfile）深度测评

对比维度	传统宿主机部署	容器化部署
环境依赖管理	需手动处理依赖冲突，如 Python 版本、PyTorch 版本、CUDA 工具链的兼容性问题，配置过程繁琐且易出错	环境隔离，每个应用拥有独立的运行环境，依赖包仅存在于容器内，避免跨应用冲突
系统清洁度	全局安装的依赖包会污染系统环境，后续清理困难，可能残留冗余文件或无效配置	依赖与配置均封装在容器中，容器删除后无残留，不影响宿主机系统清洁度
环境一致性	易出现 “在我机器上能跑” 问题，团队协作时因设备配置差异导致复现困难，排查成本高	基于镜像构建，开发、测试、生产环境完全一致，消除环境差异带来的部署隐患
迁移与部署效率	更换机器需重新执行所有配置步骤（安装依赖、调试参数等），迁移成本高、耗时久	一次构建镜像，支持 “随处运行”，新设备仅需拉取镜像即可启动，部署效率大幅提升
版本控制能力	无统一版本管理机制，依赖更新或配置修改后难以回滚到历史可用状态	镜像支持版本化标记（如 `v1.0` `v2.1`），可快速回滚到指定版本，风险可控
核心痛点总结	依赖冲突频发、系统污染、复现难、迁移成本高	无上述痛点，且额外具备资源隔离、可移植性强等优势（需掌握容器基础操作）

二、Docker Run：快速启动官方镜像

1、方案概述

这是最简单的容器化部署方式，适合快速验证和体验。

2、实战步骤

拉取官方镜像

docker pull universonic/stable-diffusion-webui:full

如果出现以下日志

root@gpu-3090-vm09:/home/sdwebui# docker pull universonic/stable-diffusion-webui:full
Using default tag: latest
Error response from daemon: Get "https://registry-1.docker.io/v2/": net/http: request canceled while waiting for connection (Client.Timeout exceeded while awaiting headers)

需要设置代理，打开vim /etc/systemd/system/docker.service.d/http-proxy.conf

vim /etc/systemd/system/docker.service.d/http-proxy.conf

填写代理信息，保存

[Service]
Environment="HTTP_PROXY=http://192.168.100.237:7890"
Environment="HTTPS_PROXY=http://192.168.100.237:7890"
Environment="NO_PROXY=localhost,127.0.0.1"

重启docker

systemctl daemon-reload
systemctl restart docker

查看代理是否生效

root@gpu-3090-vm09:/etc/systemd/system/docker.service.d# docker info

Server:
#...省略内容
 HTTP Proxy: http://192.168.100.237:7890
 HTTPS Proxy: http://192.168.100.237:7890
 No Proxy: localhost,127.0.0.1

# 创建模型和输出目录
mkdir -p ~/sd-data/models
mkdir -p ~/sd-data/outputs

将此目录挂载到容器后会继承宿主机的权限属性（UID/GID 一致），而容器内运行用户是 sduser（非 root），因此出现权限 denied。核心解决方案是 让宿主机挂载目录的 UID/GID 与容器内 sduser 的 UID/GID 完全匹配，以下是分步骤的根治方案：

# 格式：sudo chown -R 容器内sduser的UID:容器内sduser的GID 宿主机目录路径
sudo chown -R 1000:1000 ~/sd-data/models/
sudo chown -R 1000:1000 ~/sd-data/outputs/

启动容器

docker run -d \
--gpus all \
--restart unless-stopped \
-p 7860:8080 \
-e HTTP_PROXY=http://192.168.100.237:7890 \
-e HTTPS_PROXY=http://192.168.100.237:7890 \
-e NO_PROXY=localhost,127.0.0.1 \
-v /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs \
-v /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models \
--name sd-webui \
--user sduser \
universonic/stable-diffusion-webui:full

3、参数详解

参数全量写法	核心作用	详细说明
`docker run`	基础命令	Docker 的核心命令，用于「创建并启动一个新容器」（若本地无指定镜像，会自动从 Docker Hub 拉取）。
`-d`	后台运行	全称 `--detach`，让容器在后台以「守护进程模式」运行，终端不会被容器日志占用（可通过 `docker logs 容器名` 查看日志）。
`--gpus all`	GPU 资源分配	允许容器访问宿主机的所有 NVIDIA GPU（需宿主机提前安装 NVIDIA 驱动 + nvidia-docker2 插件），是 Stable Diffusion 启用 GPU 加速的核心配置（无此参数则默认用 CPU 运行，生成图片极慢）。扩展：若想指定 GPU，可写 `--gpus "device=0,1"`（仅允许容器用第 0、1 号 GPU）。
`--restart unless-stopped`	容器重启策略	配置容器退出后的重启规则，`unless-stopped` 表示： 1. 容器异常退出（如崩溃、OOM）时自动重启； 2. 手动执行 `docker stop` 停止容器后，不会自动重启； 3. 宿主机重启后，容器会随 Docker 服务自动启动（适合需要长期运行的应用，如 SD WebUI）。其他可选策略：`always`（无论何种退出都重启）、`on-failure`（仅失败退出时重启）、`no`（默认，不重启）。
`-p 7860:8080`	端口映射	全称 `--publish`，格式为「宿主机端口：容器内端口」，实现宿主机与容器的端口打通：- 左侧 `7860`：宿主机对外暴露的端口（外部访问 SD WebUI 时用这个端口，如 `http://宿主机IP:7860`）；- 右侧 `8080`：容器内 SD WebUI 实际运行的端口（需与 SD 启动命令中 `--port` 指定的端口一致，否则无法访问）。
`-e HTTP_PROXY=http://192.168.100.237:7890`	环境变量（HTTP 代理）	全称 `--env`，向容器内注入环境变量，这里用于配置 HTTP 协议的代理：当容器内进程（如 SD WebUI 下载模型、插件时）发起 HTTP 请求，会通过 `192.168.100.237:7890` 这个代理服务器转发（解决内网无法直接访问外网资源的问题）。
`-e HTTPS_PROXY=http://192.168.100.237:7890`	环境变量（HTTPS 代理）	与上一个参数类似，用于配置 HTTPS 协议的代理（注意：即使是 HTTPS 代理，这里地址也用 `http://` 开头，因为代理服务器本身的地址是 HTTP 协议）。
`-e NO_PROXY=localhost,127.0.0.1`	环境变量（无需代理的地址）	指定「不需要走代理」的地址列表：- `localhost` 和 `127.0.0.1` 均指代「容器自身」；作用：避免容器内进程访问自身服务时（如 SD 内部组件通信）也走代理，减少网络延迟和代理服务压力。
`-v /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs`	目录挂载（输出目录）	全称 `--volume`，实现「宿主机目录」与「容器内目录」的双向数据映射，核心作用是持久化数据（容器删除后数据不丢失）：- 左侧 `/home/sdwebui/sd-data/outputs`：宿主机上用于存储 SD 生成图片的目录；- 右侧 `/app/stable-diffusion-webui/outputs`：容器内 SD WebUI 默认的图片输出目录；效果：容器内生成的图片会实时同步到宿主机目录，宿主机修改该目录下的文件也会同步到容器内。
`-v /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models`	目录挂载（模型目录）	与上一个 `-v` 逻辑一致，用于挂载 SD 模型文件目录：- 宿主机目录可提前存放模型（如 Stable Diffusion v1.5、RealVis 等 `.safetensors` 或 `.ckpt` 文件）；- 容器内 SD WebUI 可直接读取该目录下的模型，无需在容器内重复下载，节省空间和时间。
`--user sduser`	指定运行用户	让容器内的进程（如 SD WebUI）以 `sduser` 这个非 root 用户身份运行：1. 安全性：避免容器内进程用 root 权限操作，降低恶意程序破坏宿主机的风险；2. 权限匹配：需确保宿主机挂载目录（如 `outputs`、`models`）的 UID/GID 与容器内 `sduser` 一致，否则会出现「权限 denied」问题（参考之前的权限配置步骤）。

4、适用场景

快速体验和测试
单次临时使用
学习Docker基础操作

5、镜像保存

如果用官方universonic/stable-diffusion-webui:full镜像每次启动都会初始化下载依赖

################################################################
Install script for stable-diffusion + Web UI
Tested on Debian 11 (Bullseye), Fedora 34+ and openSUSE Leap 15.4 or newer.
################################################################

################################################################
Running on sduser user
################################################################

################################################################
Repo already cloned, using it as install directory
################################################################

################################################################
Create and activate python venv
################################################################

################################################################
Launching launch.py...
################################################################
Python 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Installing xformers

建议等待下载完成后通过docker commit保存为新镜像

# 保存镜像
docker commit sd-webui sd-webui-new

# 运行新镜像
docker run -d \
--gpus all \
--restart unless-stopped \
-p 7860:8080 \
-v /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs \
-v /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models \
--name sd-webui \
--user sduser \
sd-webui-new

################################################################
Install script for stable-diffusion + Web UI
Tested on Debian 11 (Bullseye), Fedora 34+ and openSUSE Leap 15.4 or newer.
################################################################

################################################################
Running on sduser user
################################################################

################################################################
Repo already cloned, using it as install directory
################################################################

################################################################
Create and activate python venv
################################################################

################################################################
Launching launch.py...
################################################################
Python 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
You are up to date with the most recent release.
Launching Web UI with arguments: --update-check --xformers --listen --port 8080

三、Docker Compose：生产级编排部署

1、方案概述

使用Docker Compose可以定义复杂的多服务架构，适合生产环境部署。

2、文件内容

# 声明 Docker Compose 版本（3.2+ 支持 GPU 相关配置，与原命令兼容性匹配）
version: "3.8"

# 定义服务列表（单服务场景，服务名可自定义，此处与容器名保持一致）
services:
  # 服务名：sd-webui（与原命令 --name 对应）
  sd-webui:
    # 镜像名：sd-webui-new（与原命令最后指定的镜像一致）
    image: sd-webui-new

    # 容器运行用户：sduser（与原命令 --user sduser 对应）
    user: sduser

    # 端口映射：宿主机7860端口 → 容器8080端口（与原命令 -p 7860:8080 完全一致）
    ports:
      - "7860:8080"  # 格式："宿主机端口:容器端口"，默认支持 TCP 协议，无需额外声明

    # 目录挂载：对应原命令的两个 -v 参数，保持宿主机与容器路径一致
    volumes:
      - /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs
      - /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models

    # GPU 配置：启用所有 GPU（与原命令 --gpus all 功能一致）
    # runtime: nvidia 用于启用 NVIDIA 运行时，deploy.resources 声明 GPU 资源预留
    runtime: nvidia
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]  # 声明需要 GPU 资源，等价于 --gpus all

    # 重启策略：容器退出后自动重启（与原命令 --restart unless-stopped 完全一致）
    restart: unless-stopped

    # 可选：容器名配置（若不指定，默认生成 "项目名_sd-webui_1"，此处显式指定与原命令一致的容器名）
    container_name: sd-webui

3、操作命令

# 启动所有服务
docker compose up -d

# 查看服务状态
docker compose logs -f

# 停止服务
docker compose down

# 更新镜像并重启
docker compose pull && docker compose up -d

4、网络配置优势

反向代理（Nginx/Traefik）
监控服务（Prometheus/Grafana）
文件管理系统
数据库服务

5、适用场景

生产环境部署
多服务架构
团队协作开发
需要高可用性和监控

四、自定义Dockerfile：高度定制化部署

五、故障排除与常见问题

1、指定挂载GPU而不是全部挂载

1.1、查看容器内部GPU数量

# 进入容器内部
root@gpu-3090-vm09:/home/sdwebui# docker exec -it sd-webui bash

# 查看GPU
sduser@bb429aab044d:~/stable-diffusion-webui$ nvidia-smi
Sun Sep 28 08:34:40 2025       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.153.02             Driver Version: 570.153.02     CUDA Version: 12.8     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3090        Off |   00000000:01:00.0 Off |                  N/A |
| 36%   31C    P8             24W /  350W |    2909MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 3090        Off |   00000000:02:00.0 Off |                  N/A |
| 36%   29C    P8             29W /  350W |       3MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+

1.2、指定GPU

# 会使用第一张 RTX 3090
docker run -d \
--gpus "device=0" \
--restart unless-stopped \
-p 7860:8080 \
-v /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs \
-v /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models \
--name sd-webui \
--user sduser \
sd-webui-new

version: "3.8"

services:
  sd-webui:
    image: sd-webui-new
    user: sduser
    ports:
      - "7860:8080"
    volumes:
      - /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs
      - /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models
    runtime: nvidia
    restart: unless-stopped
    container_name: sd-webui
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              device_ids: ["0"]  # <-- 核心修改点：指定使用索引为 0 的 GPU

sduser@dec6235c8a3c:~/stable-diffusion-webui$ nvidia-smi 
Sun Sep 28 08:39:58 2025       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.153.02             Driver Version: 570.153.02     CUDA Version: 12.8     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3090        Off |   00000000:01:00.0 Off |                  N/A |
| 36%   31C    P8             24W /  350W |    2909MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+

2、网络波动，一般是因为连接超时

################################################################
Install script for stable-diffusion + Web UI
Tested on Debian 11 (Bullseye), Fedora 34+ and openSUSE Leap 15.4 or newer.
################################################################

################################################################
Running on sduser user
################################################################

################################################################
Repo already cloned, using it as install directory
################################################################

################################################################
Create and activate python venv
################################################################

################################################################
Launching launch.py...
################################################################
Python 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Installing xformers
Traceback (most recent call last):
  File "/app/stable-diffusion-webui/launch.py", line 48, in <module>
    main()
  File "/app/stable-diffusion-webui/launch.py", line 39, in main
    prepare_environment()
  File "/app/stable-diffusion-webui/modules/launch_utils.py", line 402, in prepare_environment
    run_pip(f"install -U -I --no-deps {xformers_package}", "xformers")
  File "/app/stable-diffusion-webui/modules/launch_utils.py", line 144, in run_pip
    return run(f'"{python}" -m pip {command} --prefer-binary{index_url_line}', desc=f"Installing {desc}", errdesc=f"Couldn't install {desc}", live=live)
  File "/app/stable-diffusion-webui/modules/launch_utils.py", line 116, in run
    raise RuntimeError("\n".join(error_bits))
RuntimeError: Couldn't install xformers.
Command: "/app/stable-diffusion-webui/venv/bin/python" -m pip install -U -I --no-deps xformers==0.0.23.post1 --prefer-binary
Error code: 1
stdout: Collecting xformers==0.0.23.post1
  Downloading xformers-0.0.23.post1-cp310-cp310-manylinux2014_x86_64.whl.metadata (1.0 kB)
Downloading xformers-0.0.23.post1-cp310-cp310-manylinux2014_x86_64.whl (213.0 MB)
                                            0.0/213.0 MB ? eta -:--:--

stderr: WARNING: Retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ProxyError('Cannot connect to proxy.', TimeoutError('_ssl.c:990: The handshake operation timed out'))': /simple/xformers/
  WARNING: Retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ProxyError('Cannot connect to proxy.', ConnectionResetError(104, 'Connection reset by peer'))': /packages/f4/89/ce8e936d3e64b3b565c16312dd6446d54f6e485f864130702c6b3b3cbe7c/xformers-0.0.23.post1-cp310-cp310-manylinux2014_x86_64.whl.metadata
WARNING: Connection timed out while downloading.
error: incomplete-download

× Download failed because not enough bytes were received (0 bytes/213.0 MB)
╰─> URL: https://files.pythonhosted.org/packages/f4/89/ce8e936d3e64b3b565c16312dd6446d54f6e485f864130702c6b3b3cbe7c/xformers-0.0.23.post1-cp310-cp310-manylinux2014_x86_64.whl

note: This is an issue with network connectivity, not pip.
hint: Consider using --resume-retries to enable download resumption.

~/stable-diffusion-webui

3、如何下载模型？

1、Models – Hugging Face

2、AIGate

3、Civitai: The Home of Open-Source Generative AI

Demo镜像经过commit后的总体积仅为12.2GB，包含一个预置镜像，如需要下载新模型，需要将下载的模型放入/home/sdwebui/sd-data/models指定的模型文件夹中
root@gpu-3090-vm09:/home/sdwebui/sd-data/models/Stable-diffusion# ll
total 6247844
drwxr-xr-x 2 ubuntu ubuntu       4096 Sep 28 08:50 ./
drwxr-xr-x 8 ubuntu ubuntu       4096 Sep 28 07:45 ../
-rw-r--r-- 1 root   root   2132625894 Sep 28 08:50 majicmixRealistic_v7.safetensors
-rw-r--r-- 1 ubuntu ubuntu 4265146304 Sep 28 07:44 v1-5-pruned-emaonly.safetensors
root@gpu-3090-vm09:/home/sdwebui/sd-data/models/Stable-diffusion# 

4、内存不足问题

生成图片时容器崩溃，日志显示OOM（Out of Memory）

version: "3.8"

services:
  sd-webui:
    image: sd-webui-new
    user: sduser
    ports:
      - "7860:8080"
    volumes:
      - /home/sdwebui/sd-data/outputs:/app/stable-diffusion-webui/outputs
      - /home/sdwebui/sd-data/models:/app/stable-diffusion-webui/models
    runtime: nvidia
    restart: unless-stopped
    container_name: sd-webui
    deploy:
      resources:
        limits:
          memory: "8G"  # 硬限制：容器最多只能使用 8GB 内存
        reservations:
          memory: "4G"  # 预留 4GB 内存
          devices:      # 预留 GPU 资源（与内存预留同级）
            - driver: nvidia
              capabilities: [gpu]
              device_ids: ["0"] # 指定使用索引为 0 的 GPU

5、日志查看和调试

# 查看实时日志
docker logs -f sd-webui

# 查看最近100行日志
docker logs --tail 100 sd-webui

# 进入容器调试
docker exec -it sd-webui bash

# 检查容器资源使用情况
docker stats sd-webui

6、无cuda支持

---
  0%|          | 0/20 [00:00<?, ?it/s]
*** Error completing request
*** Arguments: ('task(wan01r2cq08hz6h)', <gradio.routes.Request object at 0x7f43ce887850>, '11111', '', [], 1, 1, 7, 512, 512, False, 0.7, 2, 'Latent', 0, 0, 0, 'Use same checkpoint', 'Use same sampler', 'Use same scheduler', '', '', [], 0, 20, 'DPM++ 2M', 'Automatic', False, '', 0.8, -1, False, -1, 0, 0, 0, False, False, 'positive', 'comma', 0, False, False, 'start', '', 1, '', [], 0, '', [], 0, '', [], True, False, False, False, False, False, False, 0, False) {}
    Traceback (most recent call last):
      File "/app/modules/call_queue.py", line 74, in f
        res = list(func(*args, **kwargs))
      File "/app/modules/call_queue.py", line 53, in f
        res = func(*args, **kwargs)
      File "/app/modules/call_queue.py", line 37, in f
        res = func(*args, **kwargs)
      File "/app/modules/txt2img.py", line 109, in txt2img
        processed = processing.process_images(p)
      File "/app/modules/processing.py", line 847, in process_images
        res = process_images_inner(p)
      File "/app/modules/processing.py", line 988, in process_images_inner
        samples_ddim = p.sample(conditioning=p.c, unconditional_conditioning=p.uc, seeds=p.seeds, subseeds=p.subseeds, subseed_strength=p.subseed_strength, prompts=p.prompts)
      File "/app/modules/processing.py", line 1346, in sample
        samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=self.txt2img_image_conditioning(x))
      File "/app/modules/sd_samplers_kdiffusion.py", line 230, in sample
        samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args=self.sampler_extra_args, disable=False, callback=self.callback_state, **extra_params_kwargs))
      File "/app/modules/sd_samplers_common.py", line 272, in launch_sampling
        return func()
      File "/app/modules/sd_samplers_kdiffusion.py", line 230, in <lambda>
        samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args=self.sampler_extra_args, disable=False, callback=self.callback_state, **extra_params_kwargs))
      File "/opt/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        return func(*args, **kwargs)
      File "/app/repositories/k-diffusion/k_diffusion/sampling.py", line 594, in sample_dpmpp_2m
        denoised = model(x, sigmas[i] * s_in, **extra_args)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/modules/sd_samplers_cfg_denoiser.py", line 249, in forward
        x_out = self.inner_model(x_in, sigma_in, cond=make_condition_dict(cond_in, image_cond_in))
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/repositories/k-diffusion/k_diffusion/external.py", line 112, in forward
        eps = self.get_eps(input * c_in, self.sigma_to_t(sigma), **kwargs)
      File "/app/repositories/k-diffusion/k_diffusion/external.py", line 138, in get_eps
        return self.inner_model.apply_model(*args, **kwargs)
      File "/app/modules/sd_hijack_utils.py", line 22, in <lambda>
        setattr(resolved_obj, func_path[-1], lambda *args, **kwargs: self(*args, **kwargs))
      File "/app/modules/sd_hijack_utils.py", line 34, in __call__
        return self.__sub_func(self.__orig_func, *args, **kwargs)
      File "/app/modules/sd_hijack_unet.py", line 50, in apply_model
        result = orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs)
      File "/app/modules/sd_hijack_utils.py", line 22, in <lambda>
        setattr(resolved_obj, func_path[-1], lambda *args, **kwargs: self(*args, **kwargs))
      File "/app/modules/sd_hijack_utils.py", line 36, in __call__
        return self.__orig_func(*args, **kwargs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/models/diffusion/ddpm.py", line 858, in apply_model
        x_recon = self.model(x_noisy, t, **cond)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/models/diffusion/ddpm.py", line 1335, in forward
        out = self.diffusion_model(x, t, context=cc)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/modules/sd_unet.py", line 91, in UNetModel_forward
        return original_forward(self, x, timesteps, context, *args, **kwargs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/modules/diffusionmodules/openaimodel.py", line 797, in forward
        h = module(h, emb, context)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/modules/diffusionmodules/openaimodel.py", line 84, in forward
        x = layer(x, context)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/modules/sd_hijack_utils.py", line 22, in <lambda>
        setattr(resolved_obj, func_path[-1], lambda *args, **kwargs: self(*args, **kwargs))
      File "/app/modules/sd_hijack_utils.py", line 34, in __call__
        return self.__sub_func(self.__orig_func, *args, **kwargs)
      File "/app/modules/sd_hijack_unet.py", line 96, in spatial_transformer_forward
        x = block(x, context=context[i])
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/modules/attention.py", line 269, in forward
        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/modules/diffusionmodules/util.py", line 123, in checkpoint
        return func(*inputs)
      File "/app/repositories/stable-diffusion-stability-ai/ldm/modules/attention.py", line 272, in _forward
        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
        return self._call_impl(*args, **kwargs)
      File "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
        return forward_call(*args, **kwargs)
      File "/app/modules/sd_hijack_optimizations.py", line 497, in xformers_attention_forward
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=get_xformers_flash_attention_op(q, k, v))
      File "/opt/conda/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 223, in memory_efficient_attention
        return _memory_efficient_attention(
      File "/opt/conda/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 321, in _memory_efficient_attention
        return _memory_efficient_attention_forward(
      File "/opt/conda/lib/python3.10/site-packages/xformers/ops/fmha/__init__.py", line 337, in _memory_efficient_attention_forward
        op = _dispatch_fw(inp, False)
      File "/opt/conda/lib/python3.10/site-packages/xformers/ops/fmha/dispatch.py", line 120, in _dispatch_fw
        return _run_priority_list(
      File "/opt/conda/lib/python3.10/site-packages/xformers/ops/fmha/dispatch.py", line 63, in _run_priority_list
        raise NotImplementedError(msg)
    NotImplementedError: No operator found for `memory_efficient_attention_forward` with inputs:
         query       : shape=(2, 4096, 8, 40) (torch.float16)
         key         : shape=(2, 4096, 8, 40) (torch.float16)
         value       : shape=(2, 4096, 8, 40) (torch.float16)
         attn_bias   : <class 'NoneType'>
         p           : 0.0
    `decoderF` is not supported because:
        xFormers wasn't build with CUDA support
        attn_bias type is <class 'NoneType'>
        operator wasn't built - see `python -m xformers.info` for more info
    `flshattF@0.0.0` is not supported because:
        xFormers wasn't build with CUDA support
        operator wasn't built - see `python -m xformers.info` for more info
    `tritonflashattF` is not supported because:
        xFormers wasn't build with CUDA support
        operator wasn't built - see `python -m xformers.info` for more info
        triton is not available
        Only work on pre-MLIR triton for now
    `cutlassF` is not supported because:
        xFormers wasn't build with CUDA support
        operator wasn't built - see `python -m xformers.info` for more info
    `smallkF` is not supported because:
        max(query.shape[-1] != value.shape[-1]) > 32
        xFormers wasn't build with CUDA support
        dtype=torch.float16 (supported: {torch.float32})
        operator wasn't built - see `python -m xformers.info` for more info
        unsupported embed per head: 40

新增指定cuda兼容的xformers

# 6. 安装所有 Python 依赖
RUN pip install \
    -i https://pypi.tuna.tsinghua.edu.cn/simple \
    --trusted-host pypi.tuna.tsinghua.edu.cn \
    -r requirements_versions.txt

# 7. 【新增】安装支持 CUDA 11.8 的 xformers
# 我们使用 xformers 官方提供的预编译 wheel 文件，确保 CUDA 支持
RUN pip install --upgrade --force-reinstall \
    --index-url https://download.pytorch.org/whl/cu118 \
    --no-deps \
    xformers==0.0.22.post7

可正常生成图片

To create a public link, set `share=True` in `launch()`.
Startup time: 50.9s (prepare environment: 41.7s, import torch: 3.8s, import gradio: 1.2s, setup paths: 1.6s, initialize shared: 0.5s, other imports: 0.5s, load scripts: 0.7s, create ui: 0.7s, gradio launch: 0.1s).
6ce0161689b3853acaa03779ec93eafe75a02f4ced659bee03f50797806fa2fa
Loading weights [6ce0161689] from /app/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors
Creating model from config: /app/configs/v1-inference.yaml
/opt/conda/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
  warnings.warn(
Applying attention optimization: xformers... done.
Model loaded in 17.7s (calculate hash: 12.5s, create model: 2.8s, apply weights to model: 1.9s, calculate empty prompt: 0.2s).
100%|██████████| 20/20 [00:01<00:00, 12.15it/s]
Total progress: 100%|██████████| 20/20 [00:01<00:00, 15.26it/s]
Total progress: 100%|██████████| 20/20 [00:01<00:00, 18.42it/s]

六、结语

通过这三种容器化部署方案，您可以根据实际需求选择最适合的方式。无论是快速体验还是生产部署，容器化都能提供稳定、可复现的环境保障。

选择建议：

初学者：从Docker Run开始，快速上手

生产环境：使用Docker Compose，便于维护和扩展

定制需求：采用自定义Dockerfile，实现完全控制

容器化部署不仅解决了环境一致性问题，更为后续的扩展、监控和维护提供了坚实基础。

码道开发者社区

码道开发者社区，聚焦华为云码道 CodeArts 代码智能体，沉淀 Agent、Skill、鸿蒙开发实战内容，供开发者查阅资料、交流技术、分享工程实践

更多推荐

AI 已经能直接改设计稿了:Windows部署Penpot，接入Codex MCP实测

码道开发者社区

【AI大模型】-DeepSeek Harness 深度解析

码道开发者社区

【AI 业务流架构师】AI 业务流架构师方法论：Agentic AI 范式转移与业务编排思维

2026 年，Anthropic CEO Dario Amodei 在 Code with Claude 大会上的一个判断广为流传：一个人运营一家十亿美元级公司的概率是 70%–80%。这句话听起来像科幻小说的设定，但它背后是一个正在发生的结构性变化——AI 把"执行"的成本压到了历史低点，同时把"编排"的价值推到了前台。这意味着我们需要一套新的思考框架，来回答两个问题：一个人如何拥有一个公司的产