Commits · ddfaaf49858d0f270411bfee537897c8241ef07f · OpenDAS / FastMoE

13 May, 2021 2 commits
- gshard gate test · ddfaaf49
  Rich Ho authored May 13, 2021
  
  ddfaaf49
- add test but cannot pass · 5a0ba835
  Rick Ho authored May 13, 2021
  
  5a0ba835
27 Apr, 2021 5 commits
- gshard prune gate cuda code and test · 56cb8c15
  Rick Ho authored Apr 27, 2021
  
  56cb8c15
- compute capacity in gshard · 3e8c263c
  Rick Ho authored Apr 27, 2021
  
  3e8c263c
- balancing cuda code · 1aced6d8
  Rick Ho authored Apr 27, 2021
  
  1aced6d8
- global exchange update variable name · 6cb6bbe4
  Rick Ho authored Apr 27, 2021
  
  6cb6bbe4
- limit by threshold cuda code · bb92d30e
  Rick Ho authored Apr 27, 2021
  
  bb92d30e
26 Apr, 2021 5 commits
- fix zero gate and constant gate · 450549a4
  Rick Ho authored Apr 26, 2021
  
  450549a4
- fix stream manager · 956d3aef
  Rick Ho authored Apr 26, 2021
  
  956d3aef
- reconstruct fmoe cuda code · bb3c8966
  Rick Ho authored Apr 26, 2021
  
  bb3c8966
- add gshard and switch gate with loss · 4c90e6e8
  Rick Ho authored Apr 26, 2021
  
  4c90e6e8
- reconstruct gate structure · 82103edb
  Rick Ho authored Apr 26, 2021
  
  82103edb
25 Mar, 2021 5 commits
- Remove wierd line break in bib · 26824495
  Rick Ho authored Mar 25, 2021
  
  26824495
- Merge pull request #23 from laekov/laekov/add-citation · 46c0574b
  Jiezhong Qiu authored Mar 25, 2021
```
citation from arxiv
```
  46c0574b
- fix bibtex format · 1f5c1ce1
  Rick Ho authored Mar 25, 2021
  
  1f5c1ce1
- Merge pull request #22 from laekov/laekov/fix-tests · d81cb62f
  Jiezhong Qiu authored Mar 25, 2021
```
fix tests after updating megatron
```
  d81cb62f
- citation from arxiv · 3839d90f
  Rick Ho authored Mar 25, 2021
  
  3839d90f
23 Mar, 2021 1 commit
- fix tests after updating megatron · 6868ed2a
  Rick Ho authored Mar 24, 2021
  
  6868ed2a
22 Mar, 2021 22 commits
- Merge pull request #20 from laekov/laekov/split-megatron · cac233f3
  Rick Ho authored Mar 22, 2021
```
Split megatron python module
```
  cac233f3
- add fastmoe.megatron to install list · 9a4434f5
  Rick Ho authored Mar 22, 2021
  
  9a4434f5
- add balance to init · 2bceb76b
  Rick Ho authored Mar 22, 2021
  
  2bceb76b
- initial split · b690f245
  Rick Ho authored Mar 22, 2021
  
  b690f245
- Merge branch 'master' of github.com:laekov/fastmoe · 66fc166e
  Rick Ho authored Mar 22, 2021
  
  66fc166e
- Revise variable name · 1b7817a5
  Sengxian authored Mar 22, 2021
  
  1b7817a5
- Update issue templates · 038b31ea
  Rick Ho authored Mar 22, 2021
  
  038b31ea
- remove debug info · bedaa47d
  Jiezhong Qiu authored Mar 22, 2021
  
  bedaa47d
- format with black · 6b5a34cc
  Jiezhong Qiu authored Mar 22, 2021
  
  6b5a34cc
- relax max line length to 120 · cf9fd12a
  Jiezhong Qiu authored Mar 22, 2021
  
  cf9fd12a
- fix pylint · 68997976
  Jiezhong Qiu authored Mar 22, 2021
  
  68997976
- load/save fp32 main weight when fp16 training · 90e394bc
  Jiezhong Qiu authored Mar 22, 2021
  
  90e394bc
- fp16/fp32 optimizer has different data structure · bcfeaf3b
  Jiezhong Qiu authored Mar 17, 2021
  
  bcfeaf3b
- remove unnecessary states when dp rank>0 · 12b23ae3
  Jiezhong Qiu authored Mar 17, 2021
  
  12b23ae3
- Create Megatron.LICENSE · bc655118
  Jiezhong Qiu authored Mar 17, 2021
  
  bc655118
- load checkpoint for expert parallel · 24541d6a
  Jiezhong Qiu authored Mar 17, 2021
  
  24541d6a
- avoid saveing duplicated optim state in dp rank>0 · f5a5d31a
  Jiezhong Qiu authored Mar 17, 2021
  
  f5a5d31a
- fix pylint · c844413b
  Jiezhong Qiu authored Mar 16, 2021
  
  c844413b
- save experts separately in each data parallel rank · 49a4678c
  Jiezhong Qiu authored Mar 16, 2021
  
  49a4678c
- Merge pull request #19 from laekov/balance · 3c2a5979
  Rick Ho authored Mar 22, 2021
```
Add balance strategy
```
  3c2a5979
- Revise variable name · e028f2ec
  Sengxian authored Mar 22, 2021
  
  e028f2ec
- Merge remote-tracking branch 'origin/master' into balance_public · 69121432
  Sengxian authored Mar 22, 2021
  
  69121432