Run log · LoRA smoke test ·

Qwen3.5-4B, sequence 2048, padded

2026-11-22_smoke-train_Qwen3.5-4B_seq2048-fill

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/Qwen3.5-4B
Loaded in 4-bit
no, 16-bit
Layers offloaded to system RAM
no
Training steps
30
Sequence length
2048 tokens
LoRA rank
16
Examples padded to full length
yes
Training time
142 s
Seconds per step, averaged over every step including the first compile step
4.73 s
VRAM after load
8.47 GiB
Peak VRAM
9.12 GiB
First logged loss
0.572
Last logged loss
0.008
Adapter size on disk
100.1 MB
PyTorch
2.14.1+cu130
Hardware
RTX 4070 Ti, 12 GB

Logged loss

StepLoss
100.5715
200.06504
300.007919