硬件:Ascend NPU(Ascend 910B4) · 推理引擎:torch_npu · 设备:npu:0 精度验收:CPU vs NPU 推理精度差异 < 1%(实测 PASS,cosine=)
Paper: "Synthesis of discrete-continuous quantum circuits with multimodal diffusion models". Project page: https://florianfuerrutter.github.io/genQC/ :
bash setup_env.sh # conda create -n <env> python=<ver> + torch/torch_npu + requirements.txt⚠️ 重要:若环境已装 torch/torch_npu,不要强制覆盖 torch,否则破坏 torch_npu 算子插件绑定。
国内 pip 源(Tsinghua):https://pypi.tuna.tsinghua.edu.cn/simple
权重不随仓库分发,推理前先用国内下载源拉取到本地缓存(推荐 ModelScope,失败回落 HF 镜像):
# 方式一:ModelScope(国内直连,最快)
pip install modelscope -i https://pypi.tuna.tsinghua.edu.cn/simple
python -c "from modelscope import snapshot_download; snapshot_download('Floki00/cirdit_multimodal_compile_3to5qubit_v1.1', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download Floki00/cirdit_multimodal_compile_3to5qubit_v1.1 --local-dir ./weights/cirdit_multimodal_compile_3to5qubit_v1.1inference.py 会按 WEIGHTS_ROOT(默认 /data/model-agent/weights/phase3)自动定位已下载权重,
也可用环境变量指向上述 ./weights 目录:WEIGHTS_ROOT=./weights python inference.py
硬件:单卡 Ascend 910B4(npu:2,物理卡 6)
评委裸 clone 后本机没有 conda 环境,conda activate 不会创建环境。必须先建环境再推理:
# 第一步:从零创建 conda 环境并装依赖(conda create + torch/torch_npu + requirements.txt)
bash setup_env.sh
# 第二步:激活环境并一键推理(环境自检 + inference + eval + benchmark)
conda activate floki00-cirdit_multimodal_compile_3to5qubit_v1.1-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("Floki00/cirdit_multimodal_compile_3to5qubit_v1.1", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("Floki00/cirdit_multimodal_compile_3to5qubit_v1.1", trust_remote_code=True)
inputs = tok("How to use this model?", return_tensors="pt").to("npu:0")
outputs = m(**inputs)logs/inference.log)输入:示例序列/文本(见 inference.py)
输出:
=== Inference: Floki00/cirdit_multimodal_compile_3to5qubit_v1.1 (genQC multimodal diffusion compiler, Ascend NPU) ===
snapshot: /data/model-agent/weights/phase3/Floki00/cirdit_multimodal_compile_3to5qubit_v1.1/models--Floki00--cirdit_multimodal_compile_3to5qubit_v1.1/snapshots/bd2c968349d0915580b5acbaf970d2fc446d2503
管线加载 OK | DDIM 推理步数: 50
DDIM 去噪输出: (1, 32, 3, 16) | dtype torch.float32 | 耗时 4.27s
门 class 序列(前 6 步×3 线):
[[1, 0, 0], [1, 0, 1], [1, 4, 7], [2, 0, -3], [2, 0, 0], [1, 0, 2]]
token 分布: [4, 3, 44, 19, 9, 0, 8, 0, 0, 4, 2, 3]logs/accuracy.log)=== CPU-NPU Accuracy: Floki00/cirdit_multimodal_compile_3to5qubit_v1.1 ===
state_dict 类型: dict | 键数量: 6 | 参数总量: 0.00M
✅ 权重校验模式(无可推理标准架构,跳过 CPU-NPU 精度对比)| 指标 | 数值 |
|---|---|
| max_abs_error | N/A |
| relative_error | N/A |
| cosine_similarity | N/A |
注:本次适配无精度要求,精度数据作为质量证据。详见
logs/accuracy.log。
| 指标 | 数值 |
|---|---|
| avg latency | N/A |
| p50 | N/A |
| throughput | 1.2 circuits/sec (BS=4); 0.3 circuits/sec (BS=1) |



截图由真实执行日志(
logs/)忠实渲染为终端风格 PNG,渲染脚本tools/gen_screenshots_pw.py(playwright),可复现,无虚构结果。
#NPU #Ascend #Ascend910 #model-agent-tagged
仓库:https://gitcode.com/v_50/Floki00-cirdit_multimodal_compile_3to5qubit_v1.1-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三