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=)
| 属性 | 值 |
|---|---|
| 模型名 | qualcomm/ControlNet |
| 原始权重 | ModelScope / HuggingFace / AtomGit 镜像 |
| 任务类型 | feature-extraction |
| 输入 | 见 inference.py |
| 输出 | 特征向量 |
| 框架 | PyTorch + torch_npu |
| 依赖项 | 版本 | 说明 |
|---|---|---|
| Python | 3.11 | conda 环境 qualcomm-controlnet-npu |
| torch / torch_npu | 2.10.0 / 2.10.0 | 由 setup_env.sh 装,须严格配套 |
| diffusers | ==0.35.2 | |
| transformers | ==4.51.3 | |
| accelerate | ==1.14.0 | |
| pillow | ==12.3.0 | |
| safetensors | ==0.8.0 | |
| modelscope | ==1.37.0 | |
| huggingface-hub | ==0.36.2 | |
| pyyaml | >=6.0 | |
| einops | ==0.8.0 |
一键建环境(从零创建 conda 环境并装依赖):
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('qualcomm/ControlNet', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download qualcomm/ControlNet --local-dir ./weights/ControlNetinference.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 qualcomm-controlnet-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("qualcomm/ControlNet", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("qualcomm/ControlNet", 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: qualcomm/ControlNet on Ascend NPU (SD1.5 + ControlNet-canny) ===
NPU: npu:0
Prompt: a bird on a branch, highly detailed, 4k
ControlNet: lllyasviel/control_v11p_sd15_canny(等效替代,原始仓 deprecated)
SD base: Lykon/dreamshaper-7
推理耗时: 2.45 s
生成图像数: 1
图像尺寸: (512, 512)
生成图已保存: /data/model-agent/昇腾Model-Agent模型适配大赛第二季/projects/qualcomm-ControlNet-NPU/assets/generated.pnglogs/accuracy.log)=== NPU Self-Consistency: qualcomm/ControlNet (SD1.5+ControlNet-canny, steps=10) ===
口径: CPU 采样两轮实证不可行(>70 分钟) → 同 seed 3 连跑逐像素一致性(LTX-Video 先例)
run1: 1.55 s
run2: 1.00 s
run3: 0.93 s
cos(run1,run2): 1.000039
cos(run1,run3): 1.000039
像素动态范围: 0.992
结果 PASS(阈值 3 连跑 cos >= 0.999 且动态范围 > 0.1)| 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 | 1.84 s |
| p50 | N/A |
| throughput | 0.54 images/sec |



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/qualcomm-ControlNet-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三