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 | businesslion/scaffdiff |
| 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 businesslion-scaffdiff-npu |
| torch / torch_npu | 2.10.0 / 2.10.0 | 由 setup_env.sh 装,须严格配套 |
| addict | ==2.4.0 # sonata.model Dict 依赖 | |
| timm | ==1.0.28 # sonata.model DropPath(timm.layers) | |
| scipy | ==1.15.3 # cKDTree(knn_interpolate/chamfer) | |
| huggingface-hub | ==0.36.2 | |
| numpy | ==2.2.6 | |
| einops | ==0.8.0 | |
| safetensors | ==0.8.0 | |
| decorator | ==5.3.1 # tbe 编译器隐式依赖(约束24:NPU 图编译链 EC0010) | |
| attrs | ==26.1.0 # tbe(te) 隐式依赖链 | |
| sympy | ==1.14.0 | |
| tornado | ==6.5.8 # te 0.4.0 | |
| psutil | ==7.2.2 # opc-tool | |
| cloudpickle | ==3.1.2 | |
| ml-dtypes | ==0.6.0 |
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('businesslion/scaffdiff', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download businesslion/scaffdiff --local-dir ./weights/scaffdiffinference.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 businesslion-scaffdiff-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("businesslion/scaffdiff", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("businesslion/scaffdiff", 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: businesslion/scaffdiff (LiDAR 场景补全 diffusion, Ascend NPU) ===
[build] PTv3+denoiser 117.4M 参数
[load] 就绪 12.2s | device npu:0
[input] 部分扫描 20000 点 | scaffold 目标 20000 点 | t=200
[x0] 单步去噪 10.24s | pred_x0 (20000, 3)
[out] 有限性 True | x∈[-40.1,40.1] y∈[-40.2,40.0] z∈[-2.4,4.1]
[out] 与 noisy 目标平均偏差 0.132m(去噪生效非恒等)
device npu:0 | 端到端 22.6s
=== 完成:PTv3 条件编码 + 单步 x₀ 去噪全链跑通(Ascend NPU) ===logs/accuracy.log)
| 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 | 0.406 steps/sec(单步 x₀ 去噪平均 2.46s,20000 点/步) |



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