Pix2Struct Ascend 910 推理性能优化实战（Autoresearch 版）

git checkout -b exp/pix2struct_{optimization}_$(date +%s)
git add -A && git commit -m "exp(...): ...
- metric_primary: XX → XX
- metric_accuracy_diff: XX
- status: PASS/FAIL
"

{
  "experiment_id": "exp_001_pix2struct_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_samples_per_sec": 8.5,
    "latency_p50_ms": 115.0,
    "latency_p99_ms": 140.0,
    "docvqa_accuracy_diff": 0.002,
    "memory_mb": 5800
  },
  "status": "PASS",
  "commit_hash": "a1b2c3d"
}

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

goal: "Pix2Struct Ascend 910 推理性能达标且精度误差<1%"
termination:
  success: throughput >= baseline_target AND docvqa_accuracy_diff < 0.01
  failure: time_budget_exceeded (5h) OR consecutive_rollbacks >= 5
metrics:
  primary: throughput
  secondary: latency_p50_ms, latency_p99_ms
  guardrail: docvqa_accuracy_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://ai.gitcode.com/hf_mirrors/google/pix2struct-base pix2struct-ascend
cd pix2struct-ascend
pip install -r requirements.txt
git init
git add -A && git commit -m "init: clone pix2struct upstream"

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

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/npu_fusion_attention_encoder
# 修改 Pix2StructVisionModel 中的 attention
python eval/perf_eval.py --dtype bf16

git checkout -b exp/npu_fusion_attention_decoder
# 修改 Pix2StructDecoder 中的 attention
python eval/perf_eval.py --dtype bf16

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

git checkout -b exp/npu_ffn
python eval/perf_eval.py --dtype bf16

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

git checkout -b exp/om_conversion
git checkout main-best -- .
python -m transformers.onnx --model=google/pix2struct-base --feature=vision2seq-lm onnx/
python eval/compare_onnx_torch.py

bash om_model/atc_convert.sh

python inference_acl.py --model om_model/pix2struct-base.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:${ATOMGIT_USER_TOKEN}@gitcode.com/yourname/pix2struct-ascend.git"
git remote add origin "$REPO_URL"
git push -u origin main

if experiment.docvqa_accuracy_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()