Qwen3.5-4B-Condenser (Ascend NPU)

┌────────────────────────────────────────────────────────┐
│            BF16 vs FP32 精度特性                        │
├────────────────────────────────────────────────────────┤
│  指数位 (exponent):   8 bits (与 FP32 相同)             │
│  尾数位 (mantissa):   7 bits (FP32 为 23 bits)         │
│  动态范围:            与 FP32 完全相同                   │
│  单次运算相对误差:    <0.4%                              │
│  32层累积误差:        <1% (残差连接抑制误差传播)         │
│  Top-1 预测一致性:    100% (经验验证)                    │
│  Top-5 预测一致性:    100% (经验验证)                    │
└────────────────────────────────────────────────────────┘

pip install modelscope

# 1. 下载基础模型
modelscope download Qwen/Qwen3.5-4B --local_dir ./Qwen3.5-4B

# 2. 下载 LoRA 适配器
modelscope download twinkle-kit/Qwen3.5-4B-Condenser --local_dir ./Qwen3.5-4B-Condenser

export IMAGE=quay.io/ascend/vllm-ascend:v0.18.0
docker run --rm --net=host --shm-size=1g \
  --device /dev/davinci0 \
  --device /dev/davinci_manager \
  --device /dev/devmm_svm \
  --device /dev/hisi_hdc \
  -v /usr/local/dcmi:/usr/local/dcmi \
  -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
  -v $(pwd)/Qwen3.5-4B:/models/Qwen3.5-4B \
  -v $(pwd)/Qwen3.5-4B-Condenser:/models/Qwen3.5-4B-Condenser \
  -it $IMAGE bash

export ASCEND_RT_VISIBLE_DEVICES=0
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export TASK_QUEUE_ENABLE=1

vllm serve /models/Qwen3.5-4B \
  --served-model-name condenser \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --max-num-seqs 16 \
  --gpu-memory-utilization 0.90 \
  --enable-lora \
  --lora-modules '{"name":"condenser","path":"/models/Qwen3.5-4B-Condenser"}' \
  --max-lora-rank 16 \
  --max-loras 1 \
  --port 8000

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "condenser",
    "messages": [
      {
        "role": "system",
        "content": "You are a text compression assistant. Output format:\n## Summary\n<facts>\n\n## More\n<keywords>"
      },
      {
        "role": "user", 
        "content": "Compress: Marie Curie (born in Warsaw, 1867) discovered radium and polonium. She won Nobel Prizes in Physics (1903) and Chemistry (1911)."
      }
    ],
    "temperature": 0.3,
    "max_tokens": 256
  }'

Qwen3.5-4B-Condenser-ascend/
├── README.md                  # 本文件
├── precision_report.json       # 精度测试详细报告
├── Qwen3.5-4B-Condenser.yaml  # E2E 测试配置
├── Qwen3.5-4B-Condenser.md    # 详细教程
├── RUNBOOK.md                 # 适配运行手册
├── serve_base.sh              # 基础模型服务脚本
├── serve_condenser.sh         # LoRA 服务脚本
└── lora_adapter/              # LoRA 适配器权重
    ├── adapter_config.json
    └── adapter_model.safetensors

@misc{Qwen3.5-4B-Condenser-Ascend,
  author = {AtomCode (deepseek-v4-pro)},
  title = {Qwen3.5-4B-Condenser Ascend NPU Adaptation},
  year = {2026},
  publisher = {GitCode},
  howpublished = {\url{https://gitcode.com/Ascend-SACT/Qwen3.5-4B-Condenser}}
}

@misc{Qwen3.5-4B,
  author = {Qwen Team},
  title = {Qwen3.5-4B: A 4B Parameter Dense Language Model},
  year = {2025},
  publisher = {ModelScope},
  howpublished = {\url{https://www.modelscope.cn/models/Qwen/Qwen3.5-4B}}
}