Get Started
Home
Topics
Search
Library
Research questionHow can LoRA adapters be merged into native 4-bit microscaling checkpoints while preserving adaptation and the quantized code plane?Re-quantizing merged weights can rewrite the discrete E2M1 code plane and erase much of the learned adaptation. The resulting artifact may also depend on the exact quantization convention used during deployment.
AI
Inference Optimization
LLM Pretraining & Post-training
Machine Learning
Latest papersRecent research connected to this question, newest first.Scale-QLoRA: Code-Invariant Adapter Merging for Native 4-bit Microscaling LLMsThe source concerns NVFP4 and MXFP4 checkpoints with native per-block scales, fixed block layouts, and E2M1 code planes. It reports results across four models and four tasks, including comparisons with merge-aware QAT-LoRA, but does not establish an accuracy ranking between the approaches.research paper · Sep 3, 2026
Related questions
How can LoRA fine-tuning optimize fixed-rank updates while respecting the induced weight-matrix geometry?How can extreme low-bit LLM quantization adapt its basis across layers, calibration data, and quantizers?How can batched LLM serving handle request-specific read-write adaptation state without corrupting streams or losing throughput?How can low-bit KV caches save autoregressive decoding memory without losing long-context retrieval quality?