Author here.<p>The constraint everyone works around is that the frozen base has to fit in VRAM. But during LoRA the base is frozen — read, never written. It doesn't need to live in VRAM, it needs to arrive before the matmul that uses it. So it sits in host RAM and streams into a small pool of pre-allocated VRAM buffers, one decoder layer at a time, prefetched one ahead on a dedicated CUDA stream. Peak VRAM becomes one layer instead of the whole model.<p>Measured on an RTX 3050 Laptop (4 GB, Windows): Llama-3.1-8B in NF4 at 119.6 tok/s, 3.32 GB peak, 100% SM occupancy. Also Qwen2.5-3B with an un-quantized bf16 base at 143 tok/s in 2.15 GB, which is CUDA OOM when trained resident on the same card. Overhead is 1.43x vs resident, measured at 0.5B — the only size on this card with a valid resident baseline, and I publish that baseline so you can check the division.<p>Most of the work wasn't speed, it was correctness. Streaming fails silently: cut the autograd path and the loss still falls because the upper layers keep learning. So the bar was bit-exactness against a resident reference of the same numerics — max abs logit difference 0.0, across nine architecture families in two precisions, as a CI test rather than a one-off. That protocol caught a PEFT dispatch defect producing 0.94 logit divergence with byte-identical weights and adapters, no crash, no warning.<p>Not claiming anything above 8B — 14B NF4 needs ~7.5 GB page-locked against a measured 7.12 GB ceiling here, so I didn't run it. All numbers are Windows, so pessimistic vs Linux.<p>Measurement records, including the ones I threw away: <a href="https://github.com/MakazhanAlpamys/Soup/tree/main/benchmarks" rel="nofollow">https://github.com/MakazhanAlpamys/Soup/tree/main/benchmarks</a><p>Write-up: <a href="https://doi.org/10.5281/zenodo.21771064" rel="nofollow">https://doi.org/10.5281/zenodo.21771064</a><p>Happy to answer anything about the scheduler or the correctness protocol.