algorithmic-efficiency
algorithmic-efficiency copied to clipboard
Add workload variants
Add workload variants for the base workloads.
This is a tracking issue. 7/8 variants along with model-diff tests have been added already.
Remaining work is to:
- [x] Submit DeepSpeech workloads without OOMs.
- [ ] Run all variants e2e.
Remaining tasks.
Fix
- [x] https://github.com/mlcommons/algorithmic-efficiency/issues/654
- [x] https://github.com/mlcommons/algorithmic-efficiency/issues/662
- [x] Criteo embed init scale workload repository variant error
Rerun variants for:
- [x] criteo embed init scale
- [x] criteo resnet
- [x] imagenet_resnet_silu
- [x] imagenet_resnet_gelu
- [x] vit GLU
- [x] conformer gelu
- [x] conformer layernorm
- [x] conformer attention temp
- [x] deepspeech specaug
- [x] wmt attention temp
Pending variant reruns with correct hyperparamters for:
- [x] deepspeech_norm_and_spec_aug
- [x] deepspeech_tanh
- [x] deepspeech_no_resnet