lightning-thunder
lightning-thunder copied to clipboard
Support NeMo MegatronImagen network
🚀 Feature
NeMo's "Imagen" network implements text-to-image generation.
Initial examine:
Found 55 distinct operations, of which 46 (83.6%) are supported
Motivation
Pitch
Work items
- [x] #144
- [x] #170
- [ ] #172
- [ ] #173
- [x] #174
- [x] #175
- [x] #176
- [x] #177
- [ ] #178
- [x] Talk to NeMo team about rewriting
nan_to_num_->nan_to_num(out of place) - [x] Talk to NeMo team about rewriting
add_->add(out of place) - [ ] Test memory use of NeMo model with out-of-place
nan_to_numinstead of the in-place variant - [ ] Test memory use of NeMo model with out-of-place
addinstead of the in-place variant
cc @tfogal
Quick update: I talked to the NeMo team and Eric had the (reasonable) concern that swapping in-place for out-of-place might increase memory consumption. The onus is on me to give it a try and look at the torch memory use with both in-place and out-of-place options; if the delta is minimal or nonexistent, we could do this.
I'll edit the issue as such.