Evaluation
Pass@1 averaged over 1–32 seeds (non-thinking model), same settings for all models.
Usage
Runs on NVIDIA H100, A100, RTX Pro 6000 and B200 GPUs.
vLLM:
vllm serve qda-lab/QDA-Qwen3-Coder-30B-A3B-W4A16
SGLang:
python -m sglang.launch_server --model-path qda-lab/QDA-Qwen3-Coder-30B-A3B-W4A16
How it was made
Weights are 4-bit integers in groups of 64, quantized with GPTQ; the attention projections are 8-bit, and the group
scales were then fine-tuned by distillation. The recipe was chosen by QDA's automated search.
Reproduce
hf download qda-lab/qda-recipe-inputs --repo-type dataset --include "QDA-Qwen3-Coder-30B-A3B-W4A16/*" --local-dir inputsqda build inputs/QDA-Qwen3-Coder-30B-A3B-W4A16/qwen3-coder-30b-a3b-w4a16.yaml --set "devices=[0,1]"
Needs the QDA library (the qda command). Recipe: qwen3-coder-30b-a3b-w4a16.yaml; inputs
(calibration sample and distillation data): QDA-Qwen3-Coder-30B-A3B-W4A16/ in qda-lab/qda-recipe-inputs.
Takes about 1.4 h on 2× H100 (measured). The rebuild matches this checkpoint's KL to BF16 within 5 %; weights are not bit-identical.
License
Apache-2.0, the same as the base model.
Citation
TBD.