Train & Fine-Tune Models
Carbon-Aware Training Scheduler
Schedules training around grid carbon intensity using CodeCarbon tracking: 43% CO2 reduction with accuracy within 0.3% of baseline.
The problem
Training runs burn compute on a fixed schedule regardless of how carbon-intensive the local power grid is at that moment.
- Your training job kicks off at 2pm because that's when someone remembered to start it.
- Sustainability reporting asks for emissions numbers nobody's actually been tracking.
- A run that could've waited three hours for a cleaner grid window didn't.
What NEO built
NEO built a PyTorch pipeline that schedules training around real-time grid carbon intensity, adds gradient accumulation to cut GPU memory use, and tracks emissions with CodeCarbon.
The result
43% CO2 reduction
Cut CO2 output by 43% with accuracy within 0.3% of the uncontrolled baseline.
From the blog · 8 min
Carbon-Aware Model Training: Cutting CO2 by 43% Without Sacrificing Accuracy
NEO built a PyTorch pipeline that schedules training around grid carbon intensity and tracks emissions with CodeCarbon—43.2% CO2 reduction on MNIST with accuracy within 0.3% of baseline.
Try this in your workspace
Paste this into NEO chat to kick off the same workflow on your own data.
Schedule my PyTorch training runs around real-time grid carbon intensity, add gradient accumulation to fit on smaller GPUs, and track emissions with CodeCarbon so I can report the reduction.
Paste it in · review the plan · get the diff
Get NEO
