# Scale-QLoRA: Code-Invariant Adapter Merging for Native 4-bit Microscaling LLMs

arXiv: https://arxiv.org/abs/2609.04526 · HTML: https://mlsanity.com/paper/2609.04526/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 26 of its 35 references that are in the map.
- First posted on arXiv 3 Sep 2026.

## Builds on (most cited first, 26 of 26)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Program Synthesis with Large Language Models](https://mlsanity.com/paper/2108.07732/) (2021)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration](https://mlsanity.com/paper/2306.00978/) (2023)
- [Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task](https://mlsanity.com/paper/1809.08887/) (2018)
- [PACT: Parameterized Clipping Activation for Quantized Neural Networks](https://mlsanity.com/paper/1805.06085/) (2018)
- [Efficient Intent Detection with Dual Sentence Encoders](https://mlsanity.com/paper/2003.04807/) (2020)
- [DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale](https://mlsanity.com/paper/2201.05596/) (2022)
- [An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction](https://mlsanity.com/paper/1909.02027/) (2019)
- [LoftQ: LoRA-Fine-Tuning-Aware Quantization for Large Language Models](https://mlsanity.com/paper/2310.08659/) (2023)
- [LSQ+: Improving low-bit quantization through learnable offsets and better initialization](https://mlsanity.com/paper/2004.09576/) (2020)
- [Microscaling Data Formats for Deep Learning](https://mlsanity.com/paper/2310.10537/) (2023)
- [S-LoRA: Serving Thousands of Concurrent LoRA Adapters](https://mlsanity.com/paper/2311.03285/) (2023)
- [Additive Powers-of-Two Quantization: An Efficient Non-uniform Discretization for Neural Networks](https://mlsanity.com/paper/1909.13144/) (2019)
- [EfficientQAT: Efficient Quantization-Aware Training for Large Language Models](https://mlsanity.com/paper/2407.11062/) (2024)
- [Memory-Efficient Fine-Tuning of Compressed Large Language Models via sub-4-bit Integer Quantization](https://mlsanity.com/paper/2305.14152/) (2023)
- [LQ-LoRA: Low-rank Plus Quantized Matrix Decomposition for Efficient Language Model Finetuning](https://mlsanity.com/paper/2311.12023/) (2023)
- [Punica: Multi-Tenant LoRA Serving](https://mlsanity.com/paper/2310.18547/) (2023)
- [Optimizing Large Language Model Training Using FP4 Quantization](https://mlsanity.com/paper/2501.17116/) (2025)
- [Low-Rank Quantization-Aware Training for LLMs](https://mlsanity.com/paper/2406.06385/) (2024)
- [Accurate LoRA-Finetuning Quantization of LLMs via Information Retention](https://mlsanity.com/paper/2402.05445/) (2024)
- [L4Q: Parameter Efficient Quantization-Aware Fine-Tuning on Large Language Models](https://mlsanity.com/paper/2402.04902/) (2024)
- [ARCQuant: Boosting NVFP4 Quantization with Augmented Residual Channels for LLMs](https://mlsanity.com/paper/2601.07475/) (2026)
- [LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning](https://mlsanity.com/paper/2505.18724/) (2025)
- [Breaking the Blocks: Continuous Low-Rank Decomposed Scaling for Unified LLM Quantization and Adaptation](https://mlsanity.com/paper/2601.22716/) (2026)

## Built on by (most cited first, 1 of 1)

- [Lost in the bf16 Cast: Exporting Ternary Language Models Can Revert Most Low-Learning-Rate Code Changes](https://mlsanity.com/paper/2610.07853/) (2026)

## Abstract (arXiv)

Merging a LoRA adapter into its base model is standard deployment practice: it removes the runtime adapter's per-forward overhead and leaves a single standalone checkpoint any serving stack can load. On a native 4-bit microscaling checkpoint (NVFP4, MXFP4) that step stops being free. The merged weights must be written back through a quantizer, which re-derives the checkpoint's discrete E2M1 code plane (roughly 90% of the artifact's bytes), so the deployed artifact becomes coupled to one quantization convention, and every later code-touching event in its lifecycle can move it. Done naively the step is worse than fragile: it deletes the adaptation, by up to 39 pp, because against an already-on-grid base the reconstruction optimum is that base. Scale-QLoRA instead adapts only the native per-block scale field, trains those scales on the deployment grid, and freezes every E2M1 code. Within a fixed native format, scale grid, block layout and code plane, merging is then a bit-exact identity and the merged artifact is code-invariant. Across four models and four tasks, Scale-QLoRA and merge-aware QAT-LoRA are both accuracy-lossless, so we claim no accuracy ordering between them; they differ structurally, in that QAT-LoRA re-derives the code plane through a quantizer while Scale-QLoRA preserves it exactly. That difference is what the lifecycle prices: nearest-rounding implementations disagree by about a point on the measured task, and more extreme rule mismatches can drive the weight-space artifact to ~0%, which we report as a sensitivity bound rather than a deployment frequency. Preserving the code plane also drops the weight-space straight-through estimator from training (3.9x per step on the dense 8B model) and enables exact rollback, code-plane deduplication, and a ~125x faster scale-only task swap.