This project is the GPU version of [Pangu-alpha](https://git.openi.org.cn/PCL-Platform.Intelligence/PanGu-Alpha), please check the original project for the details of Pangu-alpha. The main purpose of this project is to enable Pangu-alpha models to be inferred and trained on GPU, so that more people can experience the charm of big models. The purpose of openness is to gather ideas and explore the potential applications of large model, as well as to identify problems that can guide our future innovative research and breakthroughs.
This project is the GPU version of [PanGu-α](https://git.openi.org.cn/PCL-Platform.Intelligence/PanGu-Alpha). Please check the original project for the details of PanGu-α. The main purpose of this project is to enable PanGu-α models to be inferred and trained on GPU such that more people can experience the charm of the big models. The purpose of openness is to gather ideas, explore the potential applications of large model, as well as identify problems that can guide our future innovative research and breakthroughs.
# mindspore Inference、Finetune、Pre-training
:
1. [Please check](inference_mindspore_gpu/README-en.md)：This part of the code only supports inference, so if you just want to experience Pangu-Alpha, we recommend using the "Three minutes to implement inference tutorial" under this page.
2. [Please check](https://gitee.com/mindspore/models/tree/master/official/nlp/pangu_alpha )：If you want to develop on Pangu-Alpha, we recommend using the training and inference code provided by mindspore. Model_zoo on the official website of mindspore provides inference, Finetune, and pre-training full process.
# MindSpore Inference, Finetune, and Pre-training
# pytorch Inference、Finetune、Pre-training
1. [Please check](inference_mindspore_gpu/README-en.md)：This part of the code only supports inference. If you just want to try PanGu-α, we recommend to use the "Three minutes to implement inference tutorial" on this page.
2. [Please check](https://gitee.com/mindspore/models/tree/master/official/nlp/pangu_alpha )：If you want to develop new techniques based on PanGu-α, we recommend to use the training and inference code provided by MindSpore. Model_zoo on the official website of MindSpore provides the full process of inference, finetune, and pre-training.
[Please check](panguAlpha_pytorch/README-en.md)：The full process of inference, Finetune, and pre-training of Pangu-Alpha developed based on Megatron-1.1.
# PyTorch Inference, Finetune, and Pre-training
[Please check](panguAlpha_pytorch/README-en.md)：The full process of inference, finetune, and pre-training of PanGu-α developed based on Megatron-1.1.
This project is a GPU inference version of [Pangu-alpha](https://git.openi.org.cn/PCL-Platform.Intelligence/PanGu-Alpha), for information about [Pangu-alpha](https://git.openi.org.cn /Intelligence/Pangu-Alpha), please see the original project for information on the principles, datasets, etc. The current phase of the project focuses on enabling Pangu-alpha models to be inferred and trained on GPUs, so that more people can experience the appeal of large models. The purpose of openness is to gather ideas, draw inspiration, and explore the potential of large model applications, as well as to identify problems that can guide our future innovative research and breakthroughs.
This project is a GPU inference version of [PanGu-α](https://git.openi.org.cn/PCL-Platform.Intelligence/PanGu-Alpha). For information about [PanGu-α](https://git.openi.org.cn /Intelligence/Pangu-Alpha), please see the original project for information on the principles, datasets, etc. The current phase of the project focuses on enabling PanGu-α models to be inferred and trained on GPUs such that more people can experience the appeal of large models. The purpose of openness is to gather ideas, draw inspiration, explore the potential of large model applications, as well as identify problems that can guide our future innovative research and breakthroughs.
@@ -12,29 +12,29 @@ This project is a GPU inference version of [Pangu-alpha](https://git.openi.org.c
[Pangu-alpha_2.6B.ckpt](https://git.openi.org.cn/attachments/27234961-4d2c-463b-9052-0240cc7ff29b?type=0) Can be used for loading 2.6B models of `fp16` and `fp32`, since the precision conversion is performed during the model loading phase
[PanGu-α_2.6B.ckpt](https://git.openi.org.cn/attachments/27234961-4d2c-463b-9052-0240cc7ff29b?type=0) can be used for loading 2.6B models of `fp16` and `fp32` since the precision conversion is performed during the model loading phase.
[ PanguAlpha_13b_fp16.ckpt](https://git.openi.org.cn/attachments/650711d6-6310-4dc2-90f8-153552e59c7a?type=0) Can only be used for loading 13B models of `fp16`
[PanGu-α_13B_fp16.ckpt](https://git.openi.org.cn/attachments/650711d6-6310-4dc2-90f8-153552e59c7a?type=0) can only be used for loading 13B models of `fp16`.
[PanguAlpha_2.6B_fp16.ckpt](https://git.openi.org.cn/attachments/7ff30c2f-e9e4-44be-8eaa-23c9d617b781?type=0) can be used for loading 2.6B models of `fp16`, is the same as [Pangu-alpha_2.6B.ckpt](https://git.openi.org.cn/attachments/27234961-4d2c-463b-9052-0240cc7ff29b?type=0), but this ckpt consumes less memory, about 20g.
[PanGu-α_2.6B_fp16.ckpt](https://git.openi.org.cn/attachments/7ff30c2f-e9e4-44be-8eaa-23c9d617b781?type=0) can be used for loading 2.6B models of `fp16`, which is the same as [PanGu-α_2.6B.ckpt](https://git.openi.org.cn/attachments/27234961-4d2c-463b-9052-0240cc7ff29b?type=0). However, this ckpt consumes less memory (about 20G).
### Graphics memory usage
| model | Graphics memory |
| Model | Graphics memory |
| --------- | --------- |
| 2.6B_fp16 | 6728 MiB |
| 2.6B_fp32 | 17214 MiB |
| 13B_fp16 | 26430 MiB |
Different models can be run depending on the video memory size of the card
Different models can be run depending on the memory size of the card.
The `2.6B_fp16` model should work on most graphics cards
The `2.6B_fp16` model should work on most graphics cards.
Already running `2.6B_fp16` model successfully on T4 and `2.6B_fp16`, `2.6B_fp32` and `13B_fp16` models on v100
The `2.6B_fp16` model has been run successfully on T4. `2.6B_fp16`, `2.6B_fp32` and `13B_fp16` models need to be run on V100 GPU.
If you don't like to use this image, you can also use the `mindspore:1.2.0` version, which allows you to run the `2.6B_fp32` model directly. There are several mindspore source code changes needed to run the `fp16` model, please see [appendix](#source code changes).
If you don't like to use this image, you can also use the `mindspore:1.2.0` version which allows you to run the `2.6B_fp32` model directly. There are several mindspore source code changes needed to run the `fp16` model. Please see [appendix](#source code changes).
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
