SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsai-for-science-oligoformer

ai-for-science-oligoformer

Agents & models

View repository

Updated 10-15 17:19

About this skill

OligoFormer Ascend NPU Migration Skill

This Skill provides complete steps for migrating OligoFormer (siRNA efficacy prediction model) from CUDA to Ascend NPU. Project address:https://github.com/lulab/OligoFormer.git

Prerequisites

Item Requirement
Hardware Ascend910 series (at least 1 card)
OS openEuler / Ubuntu / KylinOS (aarch64 or x86_64)
CANN ≥ 8.0 (8.2+ or 8.3.RC1 recommended)
Python 3.10 (original project required 3.8, upgraded to 3.10 because torch 2.5.1 does not support 3.8)
PyTorch 2.5.1
torch_npu 2.5.1
conda Installed

Migration Workflow

1. Environment initialization and CANN configuration

# 设置 CANN 环境（路径根据实际安装位置调整）
source /home/Ascend/ascend-toolkit/set_env.sh

# 验证 NPU 可用
npu-smi info

# 设置华为镜像源
export PIP_INDEX_URL=https://repo.huaweicloud.com/repository/pypi/simple/

2. Clone repository

cd /home/show
git clone https://github.com/lulab/OligoFormer.git
cd OligoFormer

3. Create conda environment and install dependencies

# 创建 Python 3.10 环境
conda create -n oligoformer-conda python=3.10 -y
conda activate oligoformer-conda

# 安装 PyTorch + torch_npu（版本需与 CANN 匹配）
pip install torch==2.5.1 torchvision==0.20.1 torch_npu==2.5.1 \
    -i https://repo.huaweicloud.com/repository/pypi/simple/

# 安装基础运行时依赖
pip install numpy==1.26.4 decorator attrs psutil absl-py cloudpickle \
    ml-dtypes scipy tornado \
    -i https://repo.huaweicloud.com/repository/pypi/simple/

# 安装项目依赖（不安装 requirements.txt 中的 torch/torchvision，已单独安装）
pip install bio matplotlib pandas prefetch-generator ptflops \
    pytorch-ignite scikit-learn tqdm yacs \
    -i https://repo.huaweicloud.com/repository/pypi/simple/

4. Install RNA-FM dependencies

OligoFormer depends on RNA-FM to generate RNA embedding features.

cd /home/show/OligoFormer

# 下载打包好的 RNA-FM（含预训练权重）
wget 'https://cloud.tsinghua.edu.cn/f/46d71884ee8848b3a958/?dl=1' -O RNA-FM.tar.gz
tar -zxf RNA-FM.tar.gz

5. Verify torch_npu availability

source /home/Ascend/ascend-toolkit/set_env.sh
python -c "import torch; import torch_npu; a = torch.randn(3,4).npu(); print(a + a)"

Should output a Tensor on device='npu:0' .

6. Code adaptation

6.1 Inject transfer_to_npu into the entry file

At the very top of scripts/main.py (before all imports), add:

import torch_npu
from torch_npu.contrib import transfer_to_npu

transfer_to_npu will automatically map torch.cuda.* to torch.npu.*, including torch.device('cuda:0') → npu:0、.cuda() → .npu() , etc.

Command-line operation:

sed -i '1i import torch_npu\nfrom torch_npu.contrib import transfer_to_npu' scripts/main.py

6.2 Add ASCEND_RT_VISIBLE_DEVICES environment variable

In the following 4 files, afterCUDA_VISIBLE_DEVICES is set, append ASCEND_RT_VISIBLE_DEVICES：

  • scripts/train.py
  • scripts/test.py
  • scripts/test_single.py
  • scripts/train_single.py
sed -i 's/os.environ\["CUDA_VISIBLE_DEVICES"\] = str(Args.cuda)/os.environ["CUDA_VISIBLE_DEVICES"] = str(Args.cuda)\n\tos.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(Args.cuda)/' \
    scripts/train.py scripts/test.py scripts/test_single.py scripts/train_single.py

# 添加 CANN 环境 source
sed -i '1a source /home/Ascend/ascend-toolkit/set_env.sh' \
    scripts/RNA-FM.sh scripts/RNA-FM-features.sh

# 替换 python 为 conda 环境完整路径
sed -i 's|^python |/root/anaconda3/envs/oligoformer-conda/bin/python |g' \
    scripts/RNA-FM.sh scripts/RNA-FM-features.sh

7. Adaptation Notes

7.1 Migration characteristics

  • Pure PyTorch project: no .cu files, no CUDA-specific third-party libraries such as flash_attn/xformers
  • No distributed: the project does not use DataParallel or DDP
  • No AMP: the project does not use mixed precision training
  • Simple CUDA usage: only used at device selection with torch.cuda.is_available() and torch.device('cuda:0')
  • transfer_to_npu can cover all CUDA APIs: no need to manually modify torch.cuda.* calls

7.2 Notes

  • model.py In PositionalEncoding in __init__ , a module-level device variable is used to create a tensor, transfer_to_npu will automatically map it to npu:0, no additional handling required
  • Ascend910 does not support fp64 (double); torch_npu will automatically downgrade to fp32, warnings may appear but results are unaffected
  • RNA-FM's setup.py contains dependency versions that are too old; version constraints need to be relaxed
  • The python command in the RNA-FM Shell script needs to point to the Python in the conda environment

8. Validation

8.1 Inference validation

source /home/Ascend/ascend-toolkit/set_env.sh
export ASCEND_RT_VISIBLE_DEVICES=0
cd /home/show/OligoFormer

python scripts/main.py --infer 1 -i1 data/example.fa

8.2 Validation pass criteria

  • Program exits normally with no NPU-related errors
  • result/ Generate RNA*_ranked.txt、RNA*_ranked_filtered.txt and other result files in the directory
  • Output contains siRNA efficacy prediction values (efficacy column is not empty)

8.3 Training validation (optional, requires generating RNA-FM features)

# 先生成训练数据的 RNA-FM 特征
bash scripts/RNA-FM-features.sh

# 训练
python scripts/main.py --datasets Hu Mix --cuda 0 --learning_rate 0.0001 \
    --batch_size 16 --epoch 200 --early_stopping 30

9. List of modified files

File Modification
scripts/main.py Inject at the top torch_npu + transfer_to_npu
scripts/train.py Add ASCEND_RT_VISIBLE_DEVICES
scripts/test.py Add ASCEND_RT_VISIBLE_DEVICES
scripts/test_single.py Add ASCEND_RT_VISIBLE_DEVICES
scripts/train_single.py Add ASCEND_RT_VISIBLE_DEVICES
RNA-FM/redevelop/launch/predict.py Inject torch_npu + Add ASCEND_RT_VISIBLE_DEVICES
RNA-FM/setup.py Relax dependency version constraints
scripts/RNA-FM.sh source CANN environment + use conda Python path
scripts/RNA-FM-features.sh source CANN environment + use conda Python path

Companion scripts

  • OligoFormer and RNA-FM precheck:python scripts/validate_oligoformer_env.py --rna-fm-path /path/to/RNA-FM

References