diff --git a/README-zh-Hans.md b/README-zh-Hans.md index 208f85671..d952a5279 100644 --- a/README-zh-Hans.md +++ b/README-zh-Hans.md @@ -34,21 +34,27 @@ Colossal-AI为您提供了一系列并行训练组件。我们的目标是让您 ## 样例 ### ViT - + + - 14倍批大小 - 5倍训练速度 -### GPT-3 & GPT-2 +### GPT-3 + + -![GPT_2_3](./docs/images/GPT_2_3.png) - GPT-3：释放 50% GPU 资源占用, 或 10.7% 加速 + +### GPT-2 + + - GPT-2：降低11倍GPU显存占用，或超线性扩展 ### BERT + -![BERT_seq](./docs/images/BERT_seq.png) - 2倍训练速度 - 1.5倍序列长度 diff --git a/README.md b/README.md index 65e05991b..1b59dda77 100644 --- a/README.md +++ b/README.md @@ -37,21 +37,27 @@ distributed training in a few lines. ## Examples ### ViT - + - 14x larger batch size - 5x faster training -### GPT-3 & GPT-2 +### GPT-3 -![GPT_2_3](./docs/images/GPT_2_3.png) + - Free 50% GPU resources, or 10.7% acceleration for GPT-3 + +### GPT-2 + + - 11x lower GPU RAM, or superlinear scaling for GPT-2 -### BERT -![BERT_seq](./docs/images/BERT_seq.png) + +### BERT + + - 2x faster training - 50% longer sequence length diff --git a/docs/images/GPT3Table.png b/docs/images/GPT3Table.png new file mode 100644 index 000000000..1754aa0b5 Binary files /dev/null and b/docs/images/GPT3Table.png differ diff --git a/docs/images/allinone/BERT_allin1.png b/docs/images/allinone/BERT_allin1.png new file mode 100644 index 000000000..c1e283ade Binary files /dev/null and b/docs/images/allinone/BERT_allin1.png differ diff --git a/docs/images/allinone/GPT2_allin1.png b/docs/images/allinone/GPT2_allin1.png new file mode 100644 index 000000000..e05e444b3 Binary files /dev/null and b/docs/images/allinone/GPT2_allin1.png differ diff --git a/docs/images/allinone/GPT3_allin1.png b/docs/images/allinone/GPT3_allin1.png new file mode 100644 index 000000000..2a96dd8c7 Binary files /dev/null and b/docs/images/allinone/GPT3_allin1.png differ diff --git a/docs/images/update/vit.png b/docs/images/update/vit.png new file mode 100644 index 000000000..37df168a4 Binary files /dev/null and b/docs/images/update/vit.png differ