In order to speed up the test, we prevent the model from generating too long thought chains, so the score may be different from that with longer thought chain.

In our experiment, the accuracy of the FP4 quantized version is almost the same as the BF16 version, and it can be used for faster inference.

Data Format	MMLU Score
BF16 Official	84.36
FP4 Quantized	80.07

Quickstart

We recommend using the Chitu inference framework(https://github.com/thu-pacman/chitu) to run this model. Here provides a simple command to show you how to run QwQ-32B-fp4.

bash

torchrun --nproc_per_node 1 \
    --master_port=22525 \
    -m chitu \
    serve.port=21002 \
    infer.cache_type=paged \
    infer.pp_size=1 \
    infer.tp_size=1 \
    models=QwQ-32B-fp4 \
    models.ckpt_dir="your model path" \
    models.tokenizer_path="your model path" \
    dtype=float16 \
    infer.do_load=True \
    infer.max_reqs=1 \
    scheduler.prefill_first.num_tasks=100 \
    infer.max_seq_len=4096 \
    request.max_new_tokens=100 \
    infer.use_cuda_graph=True

Contact

solution@qingcheng.ai

qingcheng-ai/QwQ-32B-fp4

Evaluation

Quickstart

Contact

qingcheng-ai/
QwQ-32B-fp4