参考资料
-
https://zhuanlan.zhihu.com/p/1937260131931882107
-
https://blog.csdn.net/qq_52065352/article/details/150394958
-
https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/Qwen_Image.pdf
-
https://modelscope.cn/aigc/imageGeneration?tab=advanced
环境初始化
测试环境为g5.12xlarge机器，4卡A10G，cuda版本
# nvidia-smi
Tue Nov 18 10:52:16 2025
NVIDIA-SMI 570.172.08 Driver Version: 570.172.08 CUDA Version: 12.8 # nvcc --version
Cuda compilation tools, release 12.4, V12.4.131
模型下载并上传到efs中
modelscope download --model Qwen/Qwen-Image
使用镜像环境调试
FROM public.ecr.aws/deep-learning-containers/pytorch-inference:2.6.0-gpu-py312
RUN pip config set global.index-url https://mirrors.aliyun.com/pypi/simple/
ENTRYPOINT ["sleep", "infinity"]
构建镜像
docker pull public.ecr.aws/deep-learning-containers/pytorch-inference:2.6.0-gpu-py312docker build -t qwenllm:imagev1 .
docker tag qwenllm:imagev1 xxxxxxx.dkr.ecr.cn-north-1.amazonaws.com.cn/qwenllm:imagev1
docker push xxxxxxx.dkr.ecr.cn-north-1.amazonaws.com.cn/qwenllm:imagev1
使用模板启动pod，将仓库和模型通过efs挂载到pod中
apiVersion: apps/v1
kind: Deployment
metadata:name: diffuse-qwen-imagenamespace: aitaolabels:app: diffuse
spec:replicas: 1selector:matchLabels:app: diffusetemplate:metadata:labels:app: diffusespec:serviceAccount: sa-service-account-apinodeSelector:eks.amazonaws.com/nodegroup: llm-ngcontainers:- name: diffuse-containerimage: xxxxxxx.dkr.ecr.cn-north-1.amazonaws.com.cn/qwenllm:imagev1ports:- containerPort: 8000name: http-apiresources:limits:nvidia.com/gpu: 4memory: "64Gi"cpu: "24"requests:nvidia.com/gpu: 4memory: "32Gi"cpu: "16"volumeMounts:- name: persistent-storagemountPath: /efsvolumes:- name: persistent-storagepersistentVolumeClaim:claimName: efs-claimrestartPolicy: Always
通过diffusers库推理
使用如下脚本
- device_map不设置为balanced会仅使用单卡（A10G只有20GB显存），导致oom
- 实际生成时也只使用单卡，每个step约1分钟时间较久
model_path = "/efs/models/Qwen/Qwen-Image"pipe = DiffusionPipeline.from_pretrained(model_path,torch_dtype=torch.bfloat16,device_map="balanced",
)
...
image = pipe(prompt="中国古典庭院，阳光明媚，高清写实",width=width,height=height,num_inference_steps=20,true_cfg_scale=4.0,generator=generator
).images[0]
多卡并行的官方示例中主要通过以下方式实现
- 使用
torch.multiprocessing.set_start_method('spawn', force=True)设置多进程启动方法为'spawn' - 每个GPU对应一个独立的进程(GPUWorker)，避免了Python GIL限制和CUDA上下文冲突
- 使用[MultiGPUManager](javascript:void(0))类管理所有GPU Worker进程，任务自动分配给空闲的GPU Worker
- 通过队列机制实现了简单的负载均衡
但是这个代码似乎不会对单个生成任务进行多卡并行推理。
- 每个生成任务会被分配到一个GPU上处理
- 通过任务队列([task_queue](javascript:void(0)))分发给不同的GPU Worker进程
- 每个任务只会在一个GPU上执行，而不是跨多个GPU并行执行
所以还是会出现oom
GPU 0 model initialization failed: CUDA out of memory. Tried to allocate 72.00 MiB. GPU 0 has a total capacity of 22.30 GiB of which 68.69 MiB is free. Process 3960521 has 22.23 GiB memory in use. Of the allocated memory 21.79 GiB is allocated by PyTorch, and 196.79 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
使用comfyUI
- https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/
仍旧使用上面的pod模板
# 安装 modelscope
pip install modelscope
下载fp8模型
# 下载模型文件
modelscope download --model Comfy-Org/Qwen-Image_ComfyUI split_files/diffusion_models/qwen_image_fp8_e4m3fn.safetensors
# 下载Text Encoder模型
modelscope download --model Comfy-Org/Qwen-Image_ComfyUI split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
# 下载VAE模型
modelscope download --model Comfy-Org/Qwen-Image_ComfyUI split_files/vae/qwen_image_vae.safetensors $ tree
.
├── diffusion_models
│ └── qwen_image_fp8_e4m3fn.safetensors
├── text_encoders
│ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
└── vae└── qwen_image_vae.safetensors
拷贝模型到ComfyUI的model目录下
aws s3 sync ./ s3://bucketname/Comfy-Org/Qwen-Image_ComfyUI/
启动ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI && pip install -i https://mirrors.ustc.edu.cn/pypi/simple/ -r requirements.txt
python main.py --port 8000
测试生成，大约耗时38s
- 查看资源消耗，仍旧没有使用多卡，目前看来得需要一个强大的单卡会更好（显存要够）。
- 部分模型参数被offload到cpu中，消耗了月20GB内存，所以理论上单卡的g5机型就能运行起来。
下载lora模型
/efs/ComfyUI/models/loras# modelscope download --model DiffSynth-Studio/Qwen-Image-LoRA-ArtAug-v1 model.safetensors --local_dir ./
使用4step 和8step 的lora模型能够大幅加快生成速度
https://www.modelscope.cn/models/lightx2v/Qwen-Image-Lightning/files
效果如下
后续需要封装为api可以参考文档