Congrats to @XYLi_Bruce and the team on QDA! It turns PTQ recipe design into a real closed-loop search and beats the best public recipes on multiple settings ๐
Agent loops need fast feedback, and LLM quantization doesn't have it: evaluations of reasoning models are slow, and PTQ methods are hard to port and combine.View quoted post on X
Meet QDA. A coding agent designs the quantization recipe for each model and precision, with a KL score that takes minutes and a library of 20+ stackable PTQ methods.
๐ Beats the best public recipe on 11 of 12 model-precision pairs: +0.9 / +1.4 / +2.2 pts at W4A16 / W4A8 / W3A16
๐ฅ GLM-5.3-Flash (320B) at NVFP4: 98.3% of FP8 accuracy, vs 96.9% for NVIDIA's release
๐ฐ Only $19โ75 of Claude Code tokens per search
โก Runs on stock vLLM, SGLang and llama.cpp, zero inference overhead
Blog, paper, code and checkpoints below ๐งต