MIT, NVIDIA debut QDA for LLM quantization
Researchers from MIT and NVIDIA introduced QDA, a closed-loop agentic framework that autonomously designs, evaluates, and optimizes post-training quantization recipes for LLMs. By replacing manual heuristic tuning with fast KL-divergence proxy scoring and a unified PTQ library, it boosts quantized model accuracy while cutting evaluation cycles down to minutes.
Post-training quantization has long been a bottleneck of manual tuning and fragile cross-repo porting, making QDA a notable milestone toward practical recursive self-improvement in AI systems engineering.
- –Fast proxy evaluation replaces tens of GPU-hours with minutes-long KL divergence scoring on held-out reasoning traces, paired with generation checks to filter repetition failures.
- –Outperforms leading public recipes across four reasoning models, lifting accuracy by up to 2.2 percentage points at low precisions (W4A16, W4A8, W3A16) in 64 GPU-hours.
- –Recovers 98.3% of FP8 accuracy under NVFP4 W4A4 on GLM-5.3-Flash (320B MoE), beating vendor baselines.
- –Generates checkpoints that run on standard serving kernels without runtime overhead, directly lowering production deployment costs.
- –Demonstrates how closed-loop agentic loops can handle low-level systems and compiler optimizations previously restricted to domain experts.
DISCOVERED
3h ago
2026-11-19
PUBLISHED
3h ago
2026-11-19
RELEVANCE
AUTHOR
omarsar0