docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:8.5.2-910b-ubuntu22.04-py3.11

export IMAGE=swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:8.5.2-910b-ubuntu22.04-py3.11
docker run --privileged \
  --name PP-DocBee \
  --net=host \
  --shm-size=1g \
  --device /dev/davinci_manager \
  --device /dev/devmm_svm \
  --device /dev/hisi_hdc \
  -v /usr/local/dcmi:/usr/local/dcmi \
  -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
  -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
  -v /etc/ascend_install.info:/etc/ascend_install.info \
  -v /mnt:/mnt \
  -w /workspace \
  -itd $IMAGE /bin/bash

# torch
python3 -c "import torch; print('torch:', torch.__version__)"
# torch_npu
python3 -c "
import torch, torch_npu
x = torch.randn(100, 100).to('npu:0')
y = torch.matmul(x, x.T)
print('NPU matmul OK, device:', y.device)
"
# paddle
python3 -c "import paddle; print('PaddlePaddle:', paddle.__version__)"
# transformers
python3 -c "from transformers import Qwen2VLConfig; print('transformers: OK')"
# safetensors
python3 -c "from safetensors.torch import load_file; print('safetensors: OK')"

# 模型权重默认会存在 /root/.paddlex/official_models/ 目录，此案例均以此目录为例
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-DocBee-2B_infer.tar

tar -xf PP-DocBee-2B_infer.tar

/root/.paddlex/official_models/PP-DocBee-2B/
├── model_state.pdparams      (4.2GB, 核心权重文件)
├── config.json               (4.0KB)
├── generation_config.json    (241B)
├── tokenizer_config.json     (5.0KB)
├── special_tokens_map.json   (958B)
├── added_tokens.json         (419B)
├── vocab.json                (3.6MB)
├── merges.txt                (1.6MB)
└── inference.yml             (47B)

python3 convert_bf16.py

python3 -c "
from transformers import Qwen2VLImageProcessor
Qwen2VLImageProcessor().save_pretrained('/root/.paddlex/official_models/PP-DocBee-2B-hf')
"

{
  "size": {"longest_edge": 1003520, "shortest_edge": 3136},
  "do_convert_rgb": true,
  "do_resize": true,
  "resample": 3,
  "do_rescale": true,
  "rescale_factor": 0.00392156862745098,
  "do_normalize": true,
  "patch_size": 14,
  "temporal_patch_size": 2,
  "merge_size": 2,
  "image_mean": [0.48145466, 0.4578275, 0.40821073],
  "image_std": [0.26862954, 0.26130258, 0.27577711],
  "image_processor_type": "Qwen2VLImageProcessor"
}

/root/.paddlex/official_models/PP-DocBee-2B-hf/
├── model.safetensors       (4.42GB, 转换后的 bfloat16 权重)
├── config.json                  (从原始模型复制)
├── generation_config.json       (从原始模型复制)
├── tokenizer_config.json        (从原始模型复制)
├── special_tokens_map.json      (从原始模型复制)
├── added_tokens.json            (从原始模型复制)
├── vocab.json                   (从原始模型复制)
├── merges.txt                   (从原始模型复制)
└── preprocessor_config.json     (新生成，图像处理器配置)

# 使用示例
python infer.py --device npu:2 --image-path ./medal_table.png

# 参数说明
--model-path: 模型名称，默认值："~/.paddlex/official_models/PP-DocBee-2B-hf"
--image-path: 推理图片路径，默认值："~/.paddlex/predict_input/medal_table.png"
--device: 设备id，默认值："npu:0"
--prompt: 模型提示词，默认值："识别这份表格的内容, 以markdown格式输出"
--max-new-tokens: 最大生成token数，默认值：1024
--ngram: N-gram 投机解码配置，设置为 0 时不使用，默认值：2

# python test_bf16.py --device npu:2 --image-path ./medal_table.png
Loading config...
Initializing model...
Loading bfloat16 weights...
  Re-tied lm_head.weight to embed_tokens.weight
Moving to npu:2...

=== Text-only test (greedy) ===
[transformers] Setting `pad_token_id` to `eos_token_id`:151645 for open-end generation.
  Input: The quick brown fox jumps over
  Output: The quick brown fox jumps over the lazy dog. The quick brown fox
  Time: 1.20s, Steps: 8

=== Text + Image test (greedy, max_new_tokens=1024 (n-gram=2)) ===
  Warmup (64 tokens)...
[transformers] Setting `pad_token_id` to `eos_token_id`:151645 for open-end generation.
[transformers] Setting `pad_token_id` to `eos_token_id`:151645 for open-end generation.
  Output:
system
You are a helpful assistant.
user
Please list all countries and their medal counts in a table format.
assistant
| 名次 | 国家/地区 | 金牌 | 银牌 | 铜牌 | 奖牌总数 |
|---|---|---|---|---|---|
| 1 | 中国（CHN） | 48 | 22 | 30 | 100 |
| 2 | 美国（USA） | 36 | 39 | 37 | 112 |
| 3 | 俄罗斯（RUS） | 24 | 13 | 23 | 60 |
| 4 | 英国（GBR） | 19 | 13 | 19 | 51 |
| 5 | 德国（GER） | 16 | 11 | 14 | 41 |
| 6 | 澳大利亚（AUS） | 14 | 15 | 17 | 46 |
| 7 | 韩国（KOR） | 13 | 11 | 8 | 32 |
| 8 | 日本（JPN） | 9 | 8 | 8 | 25 |
| 9 | 意大利（ITA） | 8 | 9 | 10 | 27 |
| 10 | 法国（FRA） | 7 | 16 | 20 | 43 |
| 11 | 荷兰（NED） | 7 | 5 | 4 | 16 |
| 12 | 乌克兰（UKR） | 7 | 4 | 11 | 22 |
| 13 | 肯尼亚（KEN） | 6 | 4 | 6 | 16 |
| 14 | 西班牙（ESP） | 5 | 11 | 3 | 19 |
| 15 | 牙买加（JAM） | 5 | 4 | 2 | 11 |


  Generation time: 19.02s
  Decode steps: 452
  Per-step time: 42.1ms