onnxruntime/orttraining/orttraining
Sherlock 07dc25e939
Compute global gradient norm according to 'enable_grad_norm_clip' (#5728)
* Introduce PassThrough op to wait for all gradient ready before weight update

* Compute gradient norm for fp32 runs

* Update FE UT expected value

* Respect enable_grad_norm_clip
2020-11-11 21:10:34 -08:00
..
core Compute global gradient norm according to 'enable_grad_norm_clip' (#5728) 2020-11-11 21:10:34 -08:00
models Add CPU send/recv for pipeline (#5315) 2020-11-11 12:41:39 -08:00
python Add megatron transforms for BART (#5521) 2020-11-11 16:21:36 -08:00
test Compute global gradient norm according to 'enable_grad_norm_clip' (#5728) 2020-11-11 21:10:34 -08:00
training_ops Compute global gradient norm according to 'enable_grad_norm_clip' (#5728) 2020-11-11 21:10:34 -08:00