Commits · 42c4071581a02fe05c70861e653abeeed5d3f88f · OpenDAS / Megatron-LM

14 Oct, 2022 1 commit
- Remove noop used to try to force scheduling and check for environment variable instead. · bdd97312
  Jared Casper authored Oct 14, 2022
  
  bdd97312
13 Oct, 2022 1 commit
- Adding proper test cases · 8b94a160
  Shanmugam Ramasamy authored Oct 13, 2022
  
  8b94a160
06 Oct, 2022 2 commits
- Adding some basic unit tests · b69e2195
  shanmugamr authored Oct 06, 2022
  
  b69e2195
- Setter for pipeline parallel split rank, remove print · 6defe188
  Eric Harper authored Oct 05, 2022
  
  6defe188
27 Sep, 2022 4 commits
- Support gradient accumulation fusion in fp16. · 07916bf2
  Jared Casper authored Sep 27, 2022
  
  07916bf2
- Error, not warn, if gradient_accumulation_fusion is requested but not available. · 2366716f
  Jared Casper authored Sep 27, 2022
  
  2366716f
- Correct some merge errors. · 55817ec9
  Jared Casper authored Sep 27, 2022
  
  55817ec9
- Small fixes. · 77753d0a
  Jared Casper authored Sep 27, 2022
  
  77753d0a
26 Sep, 2022 2 commits
- Check for pipeline_parallel > 2 when using interleaving. · e7e99721
  Jared Casper authored Sep 26, 2022
  
  e7e99721
- Update exports. · cbf780d0
  Jared Casper authored Sep 26, 2022
  
  cbf780d0
24 Sep, 2022 3 commits
- Improve docstrings, destory global memory buffer. · e00a1ca1
  Jared Casper authored Sep 23, 2022
  
  e00a1ca1
- Move get_num_layers into transformer.py. · a94d0a6e
  Jared Casper authored Sep 23, 2022
  
  a94d0a6e
- Alias core.parallel_state as mpu and use it throughout code. RIP mpu. · 5942af97
  Jared Casper authored Sep 23, 2022
  
  5942af97
23 Sep, 2022 3 commits
- Move layers from mpu to core.tensor_parallel. · c2ea914f
  Jared Casper authored Sep 23, 2022
  
  c2ea914f
- Bring mpu.data into megatron.core. · 209f91c9
  Jared Casper authored Sep 22, 2022
  
  209f91c9
- Start Megatron-Core with vocab parallel cross entropy · 2e6a46e4
  Jared Casper authored Sep 22, 2022
```
Also merged in some changed from apex
```
  2e6a46e4