add dsa index aoa and log

[xingmingyyj]

Before submitting

• Lint code. If there are lint issues, please format the code first.

# Install and register `pre-commit` in the project folder
pip install pre-commit && pre-commit install

# Process previous code files separately
pre-commit run --file XXXX.py

• Add test cases into tests folder. If there are codecov issues, please add tests cases first.

PR types

Others

PR changes

Others

Description

add dsa index aoa and log

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 26.08696% with 17 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@79332b4). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddleformers/trainer/trainer.py	23.07%	10 Missing ⚠️
paddleformers/transformers/aoa_config_base.py	22.22%	7 Missing ⚠️

❌ Your patch status has failed because the patch coverage (26.08%) is below the target coverage (75.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #4490   +/-   ##
==========================================
  Coverage           ?   46.44%           
==========================================
  Files              ?      475           
  Lines              ?    90639           
  Branches           ?        0           
==========================================
  Hits               ?    42098           
  Misses             ?    48541           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Paddle-CI-Bot]

PaddleFormers Log Analysis

Run #26150815783 · Attempt 1

日志分析报告

流水线名称	问题标签	修复建议
CI_ILUVATAR	硬件兼容性问题 — paddle.device.cuda.synchronize 在天数（Iluvatar）环境不可用	在 training_pipeline_step 中触发 Pipeline Parallel 的 timer 时调用了 paddle.device.cuda.synchronize()，天数为非 CUDA 设备，需将该调用替换为 paddle.device.synchronize() 或做设备类型判断跳过

---

失败的测试 case:

scripts/iluvatar_ci/test_ernie_21b_sft.py::test_ernie_21b_sft_training

---

根本原因分析:

本次 PR（分支 fix_glm5）在 paddleformers/trainer/trainer.py 中新增了 18 行代码，新增的 training_pipeline_step（trainer.py:3891）调用了 model.forward_backward_pipeline()，触发了 Fleet 内部 Pipeline Parallel 的 P2P 通信计时器（_timers("recv_forward").start()）。

该计时器内部在 timer_helper.py:62 调用了 已废弃的 paddle.device.cuda.synchronize()。天数（Iluvatar）机器使用自定义设备后端（libpaddle-iluvatar-gpu.so），Paddle 编译时未包含 CUDA 支持，导致直接抛出：

RuntimeError: (Unavailable) Paddle is not compiled with CUDA.
Cannot visit device synchronize.

---

修复建议:

在 trainer.py 的 training_pipeline_step 中，调用 model.forward_backward_pipeline() 之前，将 Fleet timer 的设备同步方式改为设备无关接口，或在进入 Pipeline Parallel 前通过 paddle.device.get_device() 判断当前设备类型，对非 CUDA 设备禁用 timer（设置 --use_timer False 或等价配置）：

# 替换方案：使用设备无关接口
paddle.device.synchronize()  # 代替 paddle.device.cuda.synchronize()

# 或在 ERNIE-21B-SFT.yaml 配置中关闭 Fleet timer（如支持）
# use_fleet_executor_timer: false

---

_{🔄 每次 Re-run 后自动更新}

[xingmingyyj]

/re-run all-failed