Commits · 2cf1d6d0a229d929e735022db295f2beea1254e3 · OpenDAS / Megatron-LM · GitLab

03 Dec, 2020 1 commit
- found a bug in consumed tokens initialization · e2a4d426
  mohammad authored Dec 02, 2020
  
  e2a4d426
02 Dec, 2020 3 commits
- addressed Jareds comments · fa80af26
  mohammad authored Dec 02, 2020
  
  fa80af26
- addressed Jareds comments · 98989693
  mohammad authored Dec 02, 2020
  
  98989693
- addrressed jareds comments · cebd3b8b
  mohammad authored Dec 02, 2020
  
  cebd3b8b
30 Nov, 2020 2 commits
- refactored learning rate scheduler so addition of variable batch size is easier · ff12df6b
  mohammad authored Nov 29, 2020
  
  ff12df6b
- added refactored learning rate · 16193619
  mohammad authored Nov 29, 2020
  
  16193619
29 Nov, 2020 2 commits
- implemented blending datasets · 65290033
  mohammad authored Nov 28, 2020
  
  65290033
- added blendable dataset · d3bb1a06
  mohammad authored Nov 28, 2020
  
  d3bb1a06
28 Nov, 2020 1 commit
- added consumed tokens to checkpoints and some refactoring · f0a445fa
  mohammad authored Nov 27, 2020
  
  f0a445fa
26 Nov, 2020 1 commit
- simplified sampler · 4311b695
  mohammad authored Nov 25, 2020
  
  4311b695
18 Nov, 2020 1 commit
- Replace deprecated torch.norm with torch.linalg.norm. · 17035d6c
  Jared Casper authored Nov 17, 2020
  
  17035d6c
17 Nov, 2020 1 commit
- Update code used for finetuning to latest API. · b219ff00
  Jared Casper authored Nov 16, 2020
  
  b219ff00
12 Nov, 2020 28 commits
- Make sure dataloader state is the same after checkpoint is loaded · cd4822f1
  Deepak Narayanan authored Nov 12, 2020
  
  cd4822f1
- Move division of loss tensor by number of microbatches to training.py · c671de3e
  Deepak Narayanan authored Nov 12, 2020
  
  c671de3e
- Small bugfix in bert_model.py: make sure word_embeddings is initialized before... · 69a546be
  Deepak Narayanan authored Nov 12, 2020
```
Small bugfix in bert_model.py: make sure word_embeddings is initialized before instantiating lm_head
```
  69a546be
- Refactor code according to Jared's comments: move pipelining and... · 1979c242
  Deepak Narayanan authored Nov 12, 2020
```
Refactor code according to Jared's comments: move pipelining and non-pipelining training loops into separate methods

Also, use mpu.get_*_model_parallel_size() instead of args.*_model_parallel_size
```
  1979c242
- Allocate tensor in `communicate()` method directly on GPU (instead of... · 9ff6f473
  mshoeybi authored Nov 11, 2020
```
Allocate tensor in `communicate()` method directly on GPU (instead of allocating on CPU and then moving to GPU)
```
  9ff6f473
- Remove timing labels that don't make sense · fdcaeba0
  Deepak Narayanan authored Nov 10, 2020
  
  fdcaeba0
- Better communication timing · 7d367b1d
  Deepak Narayanan authored Nov 10, 2020
  
  7d367b1d
- Divide gradient by number of microbatches in minibatch · 3d7194c4
  Deepak Narayanan authored Nov 03, 2020
  
  3d7194c4
- Better 'forward' and 'backward' timing in megatron/training.py · a6756bf8
  Deepak Narayanan authored Nov 03, 2020
  
  a6756bf8
- Move training schedule to 1F1B for memory efficiency · 3e6898e6
  Deepak Narayanan authored Oct 24, 2020
  
  3e6898e6
- Only transpose hidden_states when necessary · 6abf39be
  Deepak Narayanan authored Nov 03, 2020
  
  6abf39be
- Refactor word_embeddings_weight() logic into separate method, and other Mohammad comments · 57c3b364
  Deepak Narayanan authored Nov 03, 2020
  
  57c3b364
- Log times for various sub-operations in forward and backward pass in main training loop · eed0062a
  Deepak Narayanan authored Oct 30, 2020
  
  eed0062a
- Throw exception if ring_exchange is not available when pipeline_model_parallel_size > 1 · 2d8de296
  Deepak Narayanan authored Oct 30, 2020
  
  2d8de296
- Bugfix in megatron/training.py: correct global_batch_size computation · 7ce373f3
  Deepak Narayanan authored Oct 29, 2020
```
Prevents data_loader from running out of training examples
```
  7ce373f3
- Improve names of identifiers used for timing in main training loop · 9d4c735a
  Deepak Narayanan authored Oct 29, 2020
  
  9d4c735a
- Clarifications in comments and minor refactoring to make main training loop more readable · 8fb2bc8c
  Deepak Narayanan authored Oct 28, 2020
  
  8fb2bc8c
- Remove unused parameter sharing logic · 1271fd73
  Deepak Narayanan authored Oct 28, 2020
  
  1271fd73
- Bugfix in main training loop: Update master_grads only after grads are correctly accumulated · 9b558566
  Deepak Narayanan authored Oct 28, 2020
  
  9b558566
- Simplify logic in megatron/fp16/fp16.py · 767e6e92
  Deepak Narayanan authored Oct 28, 2020
  
  767e6e92
- Small notes in comments in response to Jared's comments · aa9cae27
  Deepak Narayanan authored Oct 28, 2020
  
  aa9cae27
- Address Jared's comments in README and loss_scaler.py · dd079406
  Deepak Narayanan authored Oct 27, 2020
  
  dd079406
- Improve time logging when num_microbatches_in_minibatch > 1 · 63740223
  Deepak Narayanan authored Oct 27, 2020
```
Make sure all forward and backward operations are accounted for
```
  63740223
- Back compatibility of checkpoints: use `model_parallel_size` when checking for equality of args · d5b526d5
  Deepak Narayanan authored Oct 26, 2020
  
  d5b526d5
- Refactor communication code in main training loop to helper method · 318d68c2
  Deepak Narayanan authored Oct 23, 2020
  
  318d68c2
- Back compatibility of checkpoints: don't rename model_parallel_rng_tracker · e805f0bd
  Deepak Narayanan authored Oct 22, 2020
  
  e805f0bd
- Removal of unneeded changes so that diff is smaller · 275d4e64
  Deepak Narayanan authored Oct 20, 2020
  
  275d4e64
- Intra-layer MP -> Tensor MP, Inter-layer MP -> Pipeline MP · 52a5f2f2
  Deepak Narayanan authored Oct 20, 2020
  
  52a5f2f2