PAI-PPU-V1.5.1 official image release notes
1. Image overview
The PAI-PPU official image helps you quickly use the ml.gp7vf.16.40xlarge resource (Zhenwu 810E) on PAI Serverless. This image integrates capabilities from the PPU, high-speed network, and PAI layers to deliver an out-of-the-box experience with optimal performance.
2. Notes
The PAI-PPU training image is supported only on the PAI platform, including components like DSW, DLC, and EAS.
The PAI-PPU inference image is supported only in PAI-EAS.
This official image update does not include a driver update. We recommend using driver version 1.5.0 for the PAI-PPU-V1.5.1 image.
This official image does not currently support DeepEP or sailSHMEM. A release date for this support has not been determined.
3. What's new
Supported regions now include China (Ulanqab), China (Beijing), China (Shanghai), and China (Hangzhou).
Updated the PPU SDK to version 1.5.1 (DeepEP and sailSHMEM features are not currently supported).
No updates to the EIC SDK in this release.
All image environments are updated to CUDA 12.6 + PyTorch 2.6.0.
Inference images are updated to vLLM 0.8.5 (supports V1) and SGLang 0.4.6.post1.
Includes the built-in PPU acext optimized operator library, delivering over 40% performance improvement for the DeepSeek-R1-int8 and Qwen3-235B-A22B-bf16 models. Test results are shown below:
(Tested in a lab environment. These results are for reference only and do not guarantee production performance.)
4. Image contents
4.1 Training images
pai-pg1-training-1.5.1-ubuntu | pai-pg1-training-1.5.1-ubuntu-latest | |
Environment | CUDA 12.3 + PyTorch 2.4 | CUDA 12.6 + PyTorch 2.6.0 |
Operating system | Ubuntu 22.04 | Ubuntu 24.04 |
PPU SDK version | 1.5.1 | 1.5.1 |
EIC SDK version | 1.3.7.1-v2504.3 | 1.3.7.1-v2504.3 |
Compatible CUDA version | 12.3 | 12.6 |
Python version | 3.10 | 3.12 |
PyTorch version | 2.4.0 | 2.6.0 |
OpenMMLab component versions |
|
|
Access |
4.2 Inference images
vllm:0.8.5-xpu1.5.1 | sglang:0.4.6.post1-xpu1.5.1 | |
Use case | vLLM 0.8.5, with integrated acext operator library | SGLang 0.4.6.post1, with integrated acext operator library |
Operating system | Ubuntu 24.04 | Ubuntu 24.04 |
PPU SDK version | 1.5.1 | 1.5.1 |
EIC SDK version | 1.3.7.1-v2504.3 | 1.3.7.1-v2504.3 |
Compatible CUDA version | 12.6 | 12.6 |
Python version | 3.10 | 3.10 |
PyTorch version | 2.6.0 | 2.6.0 |
Inference engine version | vLLM 0.8.5 | SGLang 0.4.6.post1 |
Validated models | DeepSeek-R1/V3, Qwen3 | DeepSeek-R1/V3, Qwen3 |
Access |
The large model inference images listed above are supported only in PAI-EAS, not in DSW or DLC. We recommend the following approach:
Use the pai-pg1-training-1.5.1-ubuntu-latest image, which is supported in DSW, DLC, and EAS. Then, install vLLM 0.8.5 or SGLang 0.4.6.post1 using pip:
pip install vllm==0.8.5+ppu1.5.1.oe pip install sglang==0.4.6.post1+ppu1.5.1.oe
5. Obtaining the images
5.1 Obtaining images directly in PAI
In your PAI workspace, navigate to the Images tab. The PAI-PPU official images are available in the PAI Official Images list.
When creating an instance in a PAI service like DSW, select an official image. You can then use the Chip Type filter to quickly locate the PAI-PPU official images.
5.2 Using inference images from Model Gallery
In Model Gallery, select a model from the DeepSeek-R1/V3 series. On the deployment page, select the single-node GP7V instance type deployment option to use the PPU-series inference images.