NPU Environment Setup

Hardware Specifications

Hardware Model	910B3 (Recommended)
Software Version	24.1.rc3
Driver Version	24.1.rc3
Firmware Version	7.8.0.2.212
Ascend CANN Toolkit Version	8.3.RC2

Environment Preparation

see https://ascend.github.io/docs/sources/ascend/quick_install.html and https://docs.vllm.ai/projects/ascend/zh-cn/v0.13.0/installation.html

# Set versions
npu="910b"
drv_version="24.1.RC3"
d_version="24.1.rc3"
firmware_ver="7.8.0.2.212"
cann_ver="8.3.RC2"
arch=$(uname -i)

# Install NPU driver
drv="Ascend-hdk-${npu}-npu-driver_${d_version}_linux-aarch64.run"
wget "https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Ascend%20HDK/Ascend%20HDK%20${drv_version}/${drv}"
bash "$drv" --full --install-for-all

# Install firmware
fw="Ascend-hdk-${npu}-npu-firmware_${firmware_ver}.run"
wget "https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Ascend%20HDK/Ascend%20HDK%20${drv_version}/${fw}"
bash "$fw" --full

# Install CANN toolkit
tk="Ascend-cann-toolkit_${cann_ver}_linux-${arch}.run"
wget --header="Referer: https://www.hiascend.com/" "https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%20${cann_ver}/${tk}"
bash "$tk" --full
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# Install kernels
kernels="Ascend-cann-kernels-${npu}_${cann_ver}_linux-${arch}.run"
wget --header="Referer: https://www.hiascend.com/" "https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%20${cann_ver}/${kernels}"
bash "$kernels" --install

# Install NNAL
nnal="Ascend-cann-nnal_${cann_ver}_linux-${arch}.run"
wget --header="Referer: https://www.hiascend.com/" "https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%20${cann_ver}/${nnal}"
bash "$nnal" --install
source /usr/local/Ascend/nnal/atb/set_env.sh

echo "Ascend NPU (${npu}) environment setup completed."

---

Key Software Versions (vLLM)

Software	Version
vllm	0.13.0
vllm-ascend	0.13.0rc1

Reference: vLLM Ascend Installation Guide A full Python environment example is provided in env.yaml.

Installation Steps

# Install vLLM
git clone --depth 1 --branch v0.13.0 https://github.com/vllm-project/vllm
cd vllm
VLLM_TARGET_DEVICE=empty pip install -v -e .
cd ..

# Install vLLM Ascend
git clone --depth 1 --branch v0.13.0rc2 https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
git submodule update --init --recursive
pip install -v -e .
cd ..

⚠️ Important: Use Python 3.10 for best compatibility.

• Step 1: Install the NPU-compatible vLLM first.

• Step 2: Install all additional Python dependencies with:

  pip install -r requirements-npu.txt

⚠️ Performance Notice: The vLLM library is the main bottleneck for NPU performance. Multi-GPU setups have not been tested and may require further optimization.

---

Running Examples

To help users quickly run the code, a sample dataset is provided:

• SFT: dataset/sft contains example bilingual data in jsonl format for SFT training.
• Monolingual: dataset/monolingual contains data needed for self-play stages.

---

Pre-download Evaluation Models

Before running evaluation, you need to prepare models for translation quality assessment.

Download the models

# Create necessary folders
mkdir -p ./models/huggingface ./models/Unbabel

# Download BLEURT-20
hf download lucadiliello/BLEURT-20 --local-dir ./models/huggingface/bleurt20

# Download COMET-Kiwi (Unbabel)
hf download Unbabel/wmt22-cometkiwi-da --local-dir ./models/Unbabel/wmt22-cometkiwi-da

Directory Structure

After downloading, the folder structure should look like this:

main.py
models/
├── huggingface/
│   └── bleurt20/
└── Unbabel/
    └── wmt22-cometkiwi-da/

⚠️ Note: Ensure the directories exist exactly as shown; otherwise evaluation scripts may fail to locate the models.

---

Quick Run Examples (Single-Card)

mkdir -p cache/qwen_debug/ ckpts/qwen_debug/
# SFT Training
torchrun \
  --nproc_per_node=1 \
  main.py \
  --mode SFT \
  --llm_path /path/to/models/Qwen3-0.6B \
  --train_data_path /path/to/data/sft \
  --cache_dir cache/qwen_debug/ \
  --output_dir ckpts/qwen_debug/ \
  --deepspeed configs/ds_z2_config.json \
  --learning_rate 3e-6 \
  --run_name qwen_baseline \
  --report_to none \
  --bf16 \
  2>&1 | tee continue.log

# RL Training
ARNOLD_WORKER_NUM=1 torchrun \
  --nproc_per_node=1 \
  main.py \
  --mode RL \
  --mcts_sample_size 1 \
  --mc_count 2 \
  --train_rounds 2 \
  --llm_path /path/to/models/Qwen3-0.6B \
  --train_data_path ./dataset/sft \
  --src_code eng_Latn \
  --trg_code zho_Hans \
  --self_play_languages "eng_Latn" "zho_Hans" "deu_Latn" \
  --cache_dir cache/qwen_debug/ \
  --flores_script flores200.py \
  --output_dir ckpts/qwen_debug/ \
  --nas_base_path . \
  --deepspeed configs/ds_z2_config.json \
  --rl_loss_type sppo_hard \
  --learning_rate 1e-4 \
  --rl_learning_rate 1e-6 \
  --run_name qwen_rl_debug \
  --report_to none \
  2>&1 | tee continue.log

An NPU run log for RL is available in continue.log.