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

git checkout -b exp/patchcore_{optimization}_$(date +%s)
git add -A && git commit -m "exp(...): ...
- metric_primary: XX → XX
- metric_auroc_diff: XX
- status: PASS/FAIL
"

{
  "experiment_id": "exp_001_patchcore_env",
  "timestamp": "2026-05-14T15:00:00Z",
  "hypothesis": "TASK_QUEUE_ENABLE=2 + CPU_AFFINITY_CONF=2 可提升推理吞吐",
  "changes": [],
  "metrics": {
    "throughput_images_per_sec": 18.0,
    "latency_p50_ms": 55.0,
    "latency_p99_ms": 72.0,
    "auroc_diff": 0.002,
    "memory_mb": 3600
  },
  "status": "PASS",
  "commit_hash": "a1b2c3d"
}

if current.throughput > best.throughput * 1.05 and current.auroc_diff < 0.01:
    git_merge_to_main()
    best = current
else:
    git_checkout_main()

goal: "Patchcore Ascend 910 推理性能达标且精度误差<1%"
termination:
  success: throughput >= baseline_target AND auroc_diff < 0.01
  failure: time_budget_exceeded (5h) OR consecutive_rollbacks >= 5
metrics:
  primary: throughput
  secondary: latency_p50_ms, latency_p99_ms
  guardrail: auroc_diff, memory_mb
loop:
  interval: "每个实验完成后立即评估"
  max_iterations: 50
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__)"

cd /workspace
git clone https://gitcode.com/gh_mirrors/pa/patchcore-inspection patchcore-ascend
cd patchcore-ascend
pip install -r requirements.txt  # torch, torchvision, scikit-learn, faiss
git init
git add -A && git commit -m "init: clone patchcore upstream"

python eval/baseline_run.py --device npu --batch 1,4,8,16

git branch main-best

git checkout -b exp/env_optimization
git checkout main-best -- .
export TASK_QUEUE_ENABLE=2
export CPU_AFFINITY_CONF=2
python eval/perf_eval.py --device npu

git checkout -b exp/backbone_optimization
# WideResNet 的 BatchNorm + ReLU 序列
# 确认 model.eval() + torch.no_grad() 已启用
# 尝试 torch.compile（如果 torch_npu 支持）
python eval/perf_eval.py

git checkout -b exp/npu_neighbor_search
# 将 CPU 上的 scikit-learn/faiss nearest neighbor 搬到 NPU
# 使用 torch.cdist 计算距离，topk 找最近邻
python eval/perf_eval.py

git checkout -b exp/contiguous_layout
python eval/perf_eval.py

git checkout -b exp/om_conversion
git checkout main-best -- .
python om_model/export_onnx.py --backbone wide_resnet50_2
python eval/compare_onnx_torch.py

bash om_model/atc_convert.sh

python inference_acl.py --model om_model/wide_resnet50_2.om --device npu
python eval/perf_eval.py --backend acl

import torch_npu
prof = torch_npu.profiler.profile(...)

python eval/final_compare.py

git tag v1.0-optimized
git checkout v1.0-optimized

mkdir -p deliverables
cp inference.py readme.md eval/ logs/ SKILL.md deliverables/

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

if experiment.auroc_diff > 0.01:
    rollback("精度超标")
elif experiment.throughput < best.throughput * 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()