GreenTune Fine-Tuning
Energy-efficient QLoRA fine-tuning on AMD MI300X with real-time power monitoring
Total Energy
0.0243kWh
87,300 Joules
Avg Power
631W
Peak: 752W / 750W TDP
Energy Cost
$0.0024
2m 19s runtime
CO2 Emissions
9.5g
0.3552 J/token
GPU Power Draw Over Time
Training Loss & Energy Efficiency
Run Configuration
Model: Qwen2.5-7B-Instruct
Quantization: 4-bit NF4
LoRA Rank: 16
LoRA Alpha: 32
Epochs: 1
Eff. Batch: 8
LR: 0.0002
Samples: 475
Datasets: hermes
Seq Length: 2048
Runtime: 2m 19s
Final Loss: 1.0544
Step-Level Metrics
| Step | Loss | Power (W) | J/Token | Tok/s | Cum. kWh | Cum. Cost | CO2 (g) | Temp (C) |
|---|---|---|---|---|---|---|---|---|
| 10 | 5.4057 | 560 | 0.0403 | 18,459 | 0.004285 | $0.0004 | 1.67 | 57 |
| 20 | 0.7899 | 596 | 0.0304 | 18,388 | 0.008218 | $0.0008 | 3.21 | 62 |
| 30 | 0.0123 | 615 | 0.0405 | 18,493 | 0.012333 | $0.0012 | 4.81 | 60 |
| 40 | 0.0050 | 625 | 0.0327 | 17,557 | 0.016361 | $0.0016 | 6.38 | 62 |
| 50 | 0.0046 | 630 | 0.0396 | 18,426 | 0.020454 | $0.0020 | 7.98 | 63 |
| 59 | 0.0000 | 631 | 0.0423 | 11,529 | 0.024250 | $0.0024 | 9.46 | 56 |
Import Training Run
Baseline (bs=2, ga=4)
Small Batch (bs=1, ga=8)