YOLOv10 Ascend 910 推理性能优化实战（Autoresearch 版）

# 每个优化尝试前自动创建分支
git checkout -b exp/conv_npu_replace_$(date +%s)
# 修改代码 → 运行 → 记录结果 → commit
git add -A && git commit -m "exp(conv): replace Conv2d with npu_conv2d

- metric_fps: 45.2 → 52.1
- metric_iou_diff: 0.003
- status: PASS
"

{
  "experiment_id": "exp_001_conv_npu",
  "timestamp": "2026-05-13T15:00:00Z",
  "hypothesis": "Conv2d → npu_conv2d 可提升卷积层性能",
  "changes": ["ultralytics/nn/modules/conv.py"],
  "metrics": {
    "fps": 52.1,
    "latency_p50_ms": 19.2,
    "latency_p99_ms": 22.1,
    "bbox_iou_diff": 0.003,
    "conf_diff": 0.001,
    "memory_mb": 2800
  },
  "status": "PASS",
  "commit_hash": "a1b2c3d"
}

# 伪代码
if current.fps > best.fps * 1.05 and current.iou_diff < 0.01:
    git_merge_to_main()
    best = current
else:
    git_checkout_main()  # 丢弃当前实验分支
    print(f"Rollback: {exp_id} 未达提升阈值或精度超标")

goal: YOLOv10 Ascend 910 推理性能在5小时内达标且精度误差<1%
termination:
  success: fps >= baseline_target AND bbox_iou_diff < 0.01 AND conf_diff < 0.01
  failure: time_budget_exceeded (5h) OR consecutive_rollbacks >= 5
metrics:
  primary: fps  # 吞吐量
  secondary: latency_p50_ms, latency_p99_ms
  guardrail: bbox_iou_diff, conf_diff, memory_mb
loop:
  interval: "每个实验完成后立即评估"
  max_iterations: 50  # 5h / 6min per exp ≈ 50 次迭代上限
actions:
  on_pass: git merge → update best → generate_next_hypothesis
  on_fail: git rollback → log_failure → try_alternative_path
  on_timeout: package_current_best → submit_deliverables

npu-smi info
python -c "import torch; import torch_npu; print(torch_npu.__version__)"

# 安装 uv（如未安装）
curl -LsSf https://astral.sh/uv/install.sh | sh

# 配置国内 PyPI 镜像（清华源）
export UV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple

# 安装依赖（示例）
uv pip install -e ./yolov10
uv pip install onnx onnxslim onnxruntime onnxscript --index-url https://pypi.tuna.tsinghua.edu.cn/simple

# HuggingFace 国内镜像
export HF_ENDPOINT=https://hf-mirror.com

# ModelScope（备用）
pip install modelscope
export MODELSCOPE_CACHE=/workspace/modelscope_cache

cd /workspace
git clone https://gitcode.com/GitHub_Trending/yo/yolov10.git
cd yolov10
pip install -e .
git init
git add -A && git commit -m "init: clone yolov10 upstream"

python eval/baseline_run.py --device npu --batch 1,8,16

git add results/baseline.json && git commit -m "baseline: NPU inference baseline

- fps_b1: XX, fps_b8: XX, fps_b16: XX
- memory_mb: XX
"

git branch main-best  # 始终保持已知最优状态

git checkout -b exp/npu_batch_nms
git checkout main-best -- .
# 修改 ultralytics/utils/ops.py
python eval/perf_eval.py --device npu
# 评估：fps > best * 1.05 且 iou_diff < 0.01 → merge，否则 rollback

git checkout -b exp/npu_conv2d
# 修改 ultralytics/nn/modules/conv.py
python eval/perf_eval.py

git checkout -b exp/npu_ciou
# 修改 ultralytics/utils/metrics.py
python eval/perf_eval.py

git checkout -b exp/contiguous_layout
# 全图扫描补 .contiguous()
python eval/perf_eval.py

git checkout -b exp/om_conversion
git checkout main-best -- .
python om_model/export_onnx.py --weights yolov10n.pt --imgsz 640
# 验证 ONNX 精度与 PyTorch 一致
python eval/compare_onnx_torch.py

bash om_model/atc_convert.sh
# 若失败 → 调整 opset / 简化图 / 修改输入 shape → 重试
# 最多 3 次尝试

python inference_acl.py --model om_model/yolov10n.om --device npu
python eval/perf_eval.py --backend acl

# 使用 ascend-profiling skill
import torch_npu
prof = torch_npu.profiler.profile(
    activities=[torch_npu.profiler.ProfilerActivity.CPU,
                torch_npu.profiler.ProfilerActivity.NPU],
    on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./Profiling_L1"),
    schedule=torch_npu.profiler.schedule(wait=0, warmup=0, active=1, repeat=1),
    experimental_config=torch_npu.profiler._ExperimentalConfig(
        aic_metrics=torch_npu.profiler.AiCMetrics.PipeUtilization,
        profiler_level=torch_npu.profiler.ProfilerLevel.Level1,
        l2_cache=False
    ),
    record_shapes=True, profile_memory=False
)

python eval/parse_profiling.py --input ./Profiling_L1 --output results/profiling_top_ops.json

python eval/final_compare.py
# 输出：Path A vs Path B 的 fps、latency、精度差异

# 从最终 commit 提取代码
mkdir -p deliverables
cp inference.py deliverables/
cp readme.md deliverables/
cp -r eval/ deliverables/
cp -r logs/ deliverables/
cp -r screenshots/ deliverables/
cp SKILL.md deliverables/

# 方式 A: 使用 ATOMGIT_USER_TOKEN
REPO_URL="https://oauth2:${ATOMGIT_USER_TOKEN}@gitcode.com/yourname/yolov10-ascend.git"
git remote add origin "$REPO_URL"
git push -u origin main

# 方式 B: 如 token 推送失败，生成 tar 包由用户手动上传

cat results/final_report.md

# PyPI 清华源
export UV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple

# HuggingFace 国内镜像
export HF_ENDPOINT=https://hf-mirror.com

# ModelScope（备用）
pip install modelscope
export MODELSCOPE_CACHE=/workspace/modelscope_cache

if experiment.accuracy_diff > 0.01:
    rollback("精度超标")
elif experiment.fps < best.fps * 0.95:
    rollback("性能倒退 > 5%")
elif experiment.memory_mb > best.memory_mb * 1.3:
    rollback("显存暴涨 > 30%")
elif experiment.latency_p99_ms > best.latency_p99_ms * 1.2:
    rollback("尾部延迟恶化 > 20%")
else:
    merge_to_main_best()