AlloyBERT 昇腾 NPU 迁移文档

docker run -it -u root -d --net=host \
--privileged \
--ipc=host \
--device=/dev/davinci_manager \
--device=/dev/devmm_svm \
--device=/dev/hisi_hdc \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/sbin:/usr/local/sbin \
-v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
--name zy-a3-ubuntu \
quay.io/ascend/cann:9.0.0-a3-ubuntu22.04-py3.11 \
/bin/bash

pip3 install torch==2.1.0 --index-url https://download.pytorch.org/whl/cpu
pip3 install torch-npu==2.1.0.post17
pip3 install transformers==4.35.2 pandas==2.1.3 scikit-learn==1.3.2 numpy==1.26.0
pip3 install pytorch-crf==0.7.2
pip3 install pyyaml tqdm matplotlib

# 检查 NPU 状态
npu-smi info

# 检查 Python 和依赖
python3 --version
python3 -c "import torch; import torch_npu; print('NPU available:', torch.npu.is_available())"

git clone https://github.com/cakshat/AlloyBERT.git /home/zy/alloybert-a3/AlloyBERT

export HF_ENDPOINT=https://hf-mirror.com
python3 -c "
from transformers import RobertaModel, RobertaTokenizerFast
model = RobertaModel.from_pretrained('roberta-base')
tokenizer = RobertaTokenizerFast.from_pretrained('roberta-base')
model.save_pretrained('/home/zy/alloybert-a3/models/roberta-base')
tokenizer.save_pretrained('/home/zy/alloybert-a3/models/roberta-base')
"

export HF_ENDPOINT=https://hf-mirror.com
python3 /home/zy/alloybert-a3/train_npu.py

export HF_ENDPOINT=https://hf-mirror.com
python3 /home/zy/alloybert-a3/inference_benchmark_npu.py

export HF_ENDPOINT=https://hf-mirror.com
python3 /home/zy/alloybert-a3/inference_demo_npu.py

# 原始代码 (来源: AlloyBERT/main.py:11)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# 适配后代码
import torch_npu
device = torch.device('npu:0')

# 1. 环境准备
export HF_ENDPOINT=https://hf-mirror.com

# 2. 克隆源码 (来源: https://github.com/cakshat/AlloyBERT)
git clone https://github.com/cakshat/AlloyBERT.git /home/zy/alloybert-a3/AlloyBERT

# 3. 下载基础模型 (来源: https://huggingface.co/roberta-base)
python3 -c "
from transformers import RobertaModel, RobertaTokenizerFast
model = RobertaModel.from_pretrained('roberta-base')
tokenizer = RobertaTokenizerFast.from_pretrained('roberta-base')
model.save_pretrained('/home/zy/alloybert-a3/models/roberta-base')
tokenizer.save_pretrained('/home/zy/alloybert-a3/models/roberta-base')
"

# 4. NPU 训练 (MPEA: MLM Pretrain + Finetune, RAYS: Finetune only, batch_size=1)
# 执行脚本: /home/zy/alloybert-a3/train_npu.py
python3 /home/zy/alloybert-a3/train_npu.py

# 5. NPU 推理 Benchmark
# 执行脚本: /home/zy/alloybert-a3/inference_benchmark_npu.py
python3 /home/zy/alloybert-a3/inference_benchmark_npu.py

# 6. NPU 推理效果演示 (可选，逐条展示推理结果)
# 执行脚本: /home/zy/alloybert-a3/inference_demo_npu.py
python3 /home/zy/alloybert-a3/inference_demo_npu.py

/home/zy/alloybert-a3/
├── train_npu.py                          # NPU 训练脚本 (MPEA: MLM Pretrain + Finetune, RAYS: Finetune only, batch_size=1)
├── inference_benchmark_npu.py            # NPU 推理 Benchmark 脚本
├── inference_demo_npu.py                 # NPU 推理效果演示脚本
├── training_results.json                 # 训练结果汇总
├── benchmark_results.json                # 推理 Benchmark 结果
├── results/                              # 推理演示输出目录
│   ├── predictions_MPEA.csv              # MPEA 逐条预测结果
│   └── predictions_RAYS.csv              # RAYS 逐条预测结果
├── models/                               # 模型文件目录
│   └── roberta-base/                     # RoBERTa-base 预训练模型 (从 HuggingFace 下载)
│       ├── config.json
│       ├── model.safetensors
│       ├── tokenizer.json
│       ├── tokenizer_config.json
│       ├── vocab.json
│       └── merges.txt
├── checkpoints/                          # 训练检查点目录
│   ├── mlm_pretrained.pt                 # MPEA MLM 预训练权重
│   ├── alloybert_MPEA_best.pt            # MPEA Pretrain+Finetune 最佳权重
│   └── alloybert_RAYS_best.pt            # RAYS Finetune 最佳权重
└── AlloyBERT/                            # 官方源码 (克隆自 https://github.com/cakshat/AlloyBERT)
    ├── main.py                           # 官方训练入口
    ├── config.yaml                       # 官方训练配置
    ├── shallowML.py                      # Shallow Models 训练脚本
    ├── requirements.txt                  # 官方依赖列表
    ├── README.md                         # 官方说明文档
    ├── LICENSE                           # MIT 许可证
    ├── MPEA_dataset.csv                  # MPEA 完整数据集 CSV
    ├── CleanMPEA.ipynb                   # MPEA 数据清洗 Jupyter Notebook
    ├── .gitignore
    ├── model/                            # 模型定义目录
    │   ├── network.py                    # AlloyBERT 模型结构定义
    │   └── utils.py                      # 训练/验证工具函数
    ├── data/                             # 数据目录
    │   ├── dataloader.py                 # 数据加载器
    │   ├── dataset.py                    # Dataset 类定义
    │   ├── convertToPKL.py               # CSV 转 PKL 脚本
    │   ├── generate_text_MPEA.py         # MPEA 文本生成脚本
    │   ├── generate_text_ysclean.py      # RAYS 文本生成脚本
    │   ├── MPEA/                         # MPEA 数据集
    │   │   ├── MPEA.csv                  # MPEA 原始数据
    │   │   ├── tr1.pkl                   # MPEA 训练集 (619 条)
    │   │   └── vl1.pkl                   # MPEA 验证集 (110 条)
    │   ├── ys_clean/                     # RAYS 数据集
    │   │   ├── ys_clean.csv              # RAYS 原始数据
    │   │   ├── tr1.pkl                   # RAYS 训练集 (691 条)
    │   │   └── vl1.pkl                   # RAYS 验证集 (122 条)
    │   ├── MPEA_clean/                   # MPEA 清洗后数据
    │   │   └── MPEA_clean.csv
    │   └── MPEA_numeric/                 # MPEA 数值化数据
    │       └── MPEA_numeric.csv
    └── tokenizer/                        # 官方自定义 BPE Tokenizer
        ├── vocab.json                    # Tokenizer 词汇表
        └── merges.txt                    # BPE 合并规则

{
  "MPEA": {
    "NPU": {
      "device": "NPU",
      "total_samples": 110,
      "total_time_s": 1.0031,
      "mse": 0.000311,
      "mae": 0.01288,
      "r2": 0.9884,
      "avg_latency_ms": 7.7,
      "p50_latency_ms": 7.67,
      "p95_latency_ms": 7.99,
      "p99_latency_ms": 8.02,
      "throughput_samples_per_sec": 109.66
    }
  },
  "RAYS": {
    "NPU": {
      "device": "NPU",
      "total_samples": 122,
      "total_time_s": 1.237,
      "mse": 0.006525,
      "mae": 0.056979,
      "r2": 0.8178,
      "avg_latency_ms": 8.23,
      "p50_latency_ms": 8.23,
      "p95_latency_ms": 8.31,
      "p99_latency_ms": 8.35,
      "throughput_samples_per_sec": 98.62
    }
  }
}

{
  "MPEA": {
    "dataset": "MPEA",
    "method": "Pretrain+Finetune",
    "best_val_mse": 0.0003109226193159887,
    "best_epoch": 36,
    "r2": 0.9884162751126911
  },
  "RAYS": {
    "dataset": "RAYS",
    "method": "Finetune only",
    "best_val_mse": 0.006525,
    "best_epoch": 31,
    "total_train_time": 1617.46,
    "r2": 0.8178
  }
}