Ankahi — Compute Log
Hardware Used
- Provider: Kaggle / Google DeepMind H100 allocation
- GPU: 1x NVIDIA H100 80GB HBM3 (MIG 3g.40gb partition)
- VRAM Available: ~40 GB
- System RAM: 128 GB
- OS: Ubuntu 22.04 LTS
Environment
- Python: 3.10.12
- PyTorch: 2.5.1+cu124
- CUDA Toolkit: 12.4
- Core Libraries: Unsloth (2026.4.6), Transformers (5.5.0), PEFT, bitsandbytes
Training Timeline & Resource Usage
| Stage | Goal | Time Elapsed | Peak VRAM |
|---|---|---|---|
| 0. Setup & Sanity | Verify pipeline | 0.5h | 11.4 GB |
| 1. Base SFT | Train multimodal adapter (Rank 16) on 16,500 pairs | 12.0h | 11.8 GB |
| 2. Persona LoRA | Train 5 separate adapters (Rank 8) on 3,000 pairs each | 10.0h (2h/persona) | 11.6 GB |
| 3. Audio Adapter | Train dysarthric audio disambiguation | 10.0h | 11.4 GB |
| 4. Safety Tuning | Train refusal behavior | 4.0h | 11.4 GB |
| 5. Merge & Deploy | Merge weights | 4.0h | 11.0 GB |
| Total | Full Pipeline Execution | ~40.5h | 11.8 GB max |
Notes on Efficiency
Thanks to Unsloth and Triton kernels, the training of Gemma 4 E4B was executed within the VRAM constraints of an H100 MIG partition without triggering OOM (Out-of-Memory) errors, provided aggressive garbage collection (gc.collect()) was enforced between persona training loops.