Skip to content

Commit

Permalink
glm: add 2x8 statistics (#216)
Browse files Browse the repository at this point in the history
Co-authored-by: zhouyu <zhouyu@baai.ac.cn>
  • Loading branch information
yuzhou03 and zhouyu authored Aug 29, 2023
1 parent 26dde69 commit 269c1fd
Showing 1 changed file with 2 additions and 1 deletion.
3 changes: 2 additions & 1 deletion training/nvidia/glm-pytorch/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -46,5 +46,6 @@
| A100单机8卡(1x8) | fp32 | / | 2763 | 36.5 | 42.4 | 42.4 | 0.808 | 33.0/40.0 | 0.035 |
| A100单机8卡(1x8) | fp32 | bs=16, lr=1e-05 | 2688 | 37.4 | 43.5 | 43.5 | 0.801 | 39.5/40.0 | 0.035 |
| A100单机单卡(1x1) | fp32 | bs=16, lr=1e-05 | 7695 | 4.2 | 5.5 | 5.5 | | 35.0/40.0 | 0.036 |
| A100两机16卡(2x8) | fp32 | bs=8, lr=1e-05 | 1409 | 72.7 | 84.7 | 84.9 | 0.804 | 31.0/40.0 | 0.036 |

> 注:使用GLMForMultiTokenCloze进行forward计算你,得到MFU=0.04, 使用GLMModel模型forward计算,得到MFU=0.08. 本模型的MFU值偏低是由于原始模型的MFU较低。
> 注:使用GLMForMultiTokenCloze进行forward计算,得到MFU=0.04, 使用GLMModel模型forward计算,得到MFU=0.08. 本模型的MFU值偏低是由于原始模型的MFU较低。

0 comments on commit 269c1fd

Please sign in to comment.