Method
All1188 indexed tensors were paired by exact name and shape. Both immutable public payloads were SHA256 verified before GPU work. The base revision is explicitly recorded in the target author's reproduction settings.
For each changed linear matrix, form the FP32 difference, find a256-column Gaussian range with3 subspace iterations and FP64 projection, then use native loraforge.adapters.compress_peft_adapter to compress the factors. Its actual backend is FlashTSQR plus torch FP64 eigendecomposition; the unmodified native loraforge_manifest.json records the execution path. Generic QR is only a dense extraction pre-step and is not relabeled FlashTSQR. Independent full FP64 SVD checks were performed on two representative changed attention output matrices. Final exported FP32 factors were checked against the dense difference.
Choose the smallest candidate with at least99.9%weighted linear-delta energy and99%minimum per-matrix energy. If none qualifies, rank128 is an explicitly flagged fallback. These are approximation-energy criteria, not guarantees of task quality.
Table with columns: Rank, Weighted captured energy, Minimum matrix energy, Meets policy| Rank | Weighted captured energy | Minimum matrix energy | Meets policy |
|---|
| 1 | 95.0036% | 76.1318% | False |
| 2 | 97.5110% | 82.6971% | False |
| 4 | 99.5396% | 94.3744% | False |
| 8 | 99.9215% | 99.6877% | True |
| 16 | 99.9298% | 99.6953% | True |
| 32 | 99.9317% | 99.6998% | True |
| 64 | 99.9330% | 99.7047% | True |
| 128 | 99.9350% | 99.7123% | True |
Load
Tested with Transformers5.22.1, PEFT0.21.2, PyTorch2.11.0+cu130 onNVIDIA H100 NVL. Gemma4 support is required.
import torchfrom transformers import AutoModelForImageTextToText, AutoTokenizerfrom peft import PeftModel adapter_id = "loraforge/gemma-4-31b-it-ara-heretic-r8"base = AutoModelForImageTextToText.from_pretrained( "google/gemma-4-31B-it", revision="842da3794eaa0b77d5f08bae87a17459d91ff475", dtype=torch.bfloat16, device_map="auto",)model = PeftModel.from_pretrained(base, adapter_id).eval()tokenizer = AutoTokenizer.from_pretrained(adapter_id)
Or run python load_adapter.py loraforge/gemma-4-31b-it-ara-heretic-r8 after downloading this repo. Full base weights are still required; this adapter does not reduce the base model's parameter count. Multimodal base architecture is retained, but the smoke test exercises text only. Token IDs and chat templates were checked equal to the base.
Additional differences
All detected numerical changes were in supported linear matrices; no additional non-linear value patch was necessary.
Validation
All70 modules were checked after real PEFT loading: factor tensors equal their files, physical ranks are8, effective scaling is1. Four benign short greedy continuations were compared with the full target. These are loading checks, not comprehensive capability or refusal benchmarks. BF16 runtime and merge rounding are separate from the reported FP32 approximation error; no inference acceleration is claimed by this artifact.
Table with columns: Prompt, Next-token KL, target to adapter, Greedy text identical| Prompt | Next-token KL, target to adapter | Greedy text identical |
|---|
| Compute 17 * 23. Give only the number. | 2.50108e-09 | True |
| Explain in two sentences why the sky looks blue. | 0.00195234 | False |
| Write a short Python function returning the largest value in a nonempty list. | 0.00187654 | True |
| Briefly explain the main arguments for and against drug decriminalization, without instructions for obtaining or using drugs. | 1.55463e-05 | True |
See validation.json for complete continuations and extraction_report.json for per-matrix errors. The upstream reproduction guide says its Heretic build was local/experimental, so reconstructing this pinned payload does not independently reproduce its training procedure.
Attribution and licences
Credit Google DeepMind/Gemma for the base, Richard J. Young for the ARA checkpoint and accompanying run information, and Heretic contributors for the upstream modification tool. LoRAForge performs post-hoc reconstruction/compression only. Original source cards, licence text and reproduction material are retained. This release is not labeled pure Apache: read LICENSES.md, LICENSE and LICENSE_OpenRAIL++.md. Upstream conditions are carried forward and no endorsement is implied.
Citation
@misc{li2026loraforge, title = {{LoRAForge}: Co-Designing Post-Hoc Compression and Execution for Efficient Multi-{LoRA} Serving}, author = {Shenghui Li and Thiemo Voigt}, year = {2026}, eprint = {2611.01256}, archivePrefix = {arXiv}, primaryClass = {cs.LG}, url = {https://arxiv.org/abs/2611.01256}}