Run log · LoRA smoke test ·

gemma-4-E4B-it, sequence 512

2026-11-22_smoke-train_gemma-4-E4B-it_seq512

A smoke test checks that training runs on this card, how fast, and with how much memory. It uses eight placeholder examples. The adapter is thrown away, and the falling loss only shows the pipeline can memorise them.

Recorded fields

Model
unsloth/gemma-4-E4B-it
Loaded in 4-bit
yes
Layers offloaded to system RAM
yes
Training steps
100
Sequence length
512 tokens
LoRA rank
16
Training time
231 s
Seconds per step, averaged over every step including the first compile step
2.31 s
VRAM after load
4.93 GiB
Peak VRAM
7.09 GiB
First logged loss
3.966
Last logged loss
0.025
Adapter size on disk
163.8 MB
PyTorch
2.14.1+cu130
Hardware
RTX 4070 Ti, 12 GB

Logged loss

StepLoss
103.966
201.481
300.3008
400.03557
500.03132
600.02572
700.02589
800.02422
900.02493
1000.02467