Nanochat RTX 4070 Training Run Update | Generated by AI
Nanochat is running on the RTX 4070:
GPU: RTX 4070 — 2.7 GB / 12 GB VRAM used, 99% utilization Model: depth=8 (~40M params), bf16 Throughput: ~103K tok/sec Loss: 7.73 → 6.22 (step 81/5000) ETA: ~26 min pretraining, then SFT (~8 min), total ~35-40 min
Running in tmux session nanochat on the workstation. Log at /mnt/data/nanochat/run_rtx4070.log.
Note: only 2.7 GB VRAM used — you could easily bump to --depth=12 or --depth=16 for a bigger model if you want. The d8 is a quick demo run. Want me to let this finish, or restart with a larger model?