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 SKIP,cosine=)
especially in multi-image and video scenarios.
| 属性 | 值 |
|---|---|
| 模型名 | openbmb/MiniCPM-V-4.6 |
| 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 |
| 官方实现 | https://github.com/OpenBMB/MiniCPM-o |
官方实现 / 论文仓库:
| 依赖项 | 版本 | 说明 |
|---|---|---|
| Python | 3.11 | conda 环境 openbmb-minicpm-v-4.6-npu |
| torch / torch_npu | 2.10.0 / 2.10.0 | 由 setup_env.sh 装,须严格配套 |
| transformers | ==4.51.3 | |
| accelerate | ==1.14.0 | |
| pillow | ==12.3.0 | |
| numpy | ==1.26.4 | |
| huggingface-hub | ==0.36.2 | |
| requests | ==2.34.2 | |
| modelscope | ==1.37.0 | |
| safetensors | ==0.8.0 | |
| multimolecule | ==0.2.1 | |
| pyyaml | >=6.0 | |
| einops | ==0.8.0 # transformers 4.51 modeling 隐式依赖(torch_npu 存在时触发) | |
| torchvision | ==0.25.0 |
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('openbmb/MiniCPM-V-4.6', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download openbmb/MiniCPM-V-4.6 --local-dir ./weights/MiniCPM-V-4.6inference.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:3,物理卡 7)
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 openbmb-minicpm-v-4.6-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("openbmb/MiniCPM-V-4.6", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("openbmb/MiniCPM-V-4.6", 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: openbmb/MiniCPM-V-4.6 on Ascend NPU ===
NPU: npu:0
权重目录: /data/model-agent/weights/phase3/openbmb/MiniCPM-V-4.6
pipeline_tag: | library: | is_text=False
state_dict 类型: <class 'dict'>
键数量: 779
model.language_model.embed_tokens.weight: (248094, 1024)
model.language_model.layers.0.input_layernorm.weight: (1024,)
model.language_model.layers.0.linear_attn.A_log: (16,)
model.language_model.layers.0.linear_attn.conv1d.weight: (6144, 1, 4)
model.language_model.layers.0.linear_attn.dt_bias: (16,)
输出 shape: (254048256,)
输出 L2 norm: 274.7086
输出前 8 维: [0.01031494140625, 0.013916015625, -0.021484375, -0.0034637451171875, 0.0004119873046875, 0.036376953125, -0.03857421875, 0.0208740234375]logs/accuracy.log)=== CPU-NPU Accuracy: openbmb/MiniCPM-V-4.6 ===
state_dict 类型: dict | 键数量: 779 | 参数总量: 1300.43M
✅ 权重校验模式(无可推理标准架构,跳过 CPU-NPU 精度对比)精度对比:SKIP(state_dict_only_no_architecture)
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 | N/A |



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/openbmb-MiniCPM-V-4.6-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三