[Benchmark] Add GLM-4.5-Air 128K PT config

[liym27]

Add benchmark config for GLM-4.5-Air (106B-A12B MoE, 46 layers) at 128K context length under PaddleFleet PT stage.

Parallelism: TP=8, PP=4, EP=8, DP=1 (32 cards)
PP layer distribution: [10, 12, 12, 12]
(num_empty_layers_add_in_head=2, num_empty_layers_add_in_tail=0)
Sequence length: 131072
Global batch size: 128
Learning rate: 5.0e-5 with cosine decay
Router aux loss: 1.0e-4
Checkpoint format: flex_checkpoint

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

PR changes

Description

[paddle-bot]

Thanks for your contribution!

[Paddle-CI-Bot]

PaddleFormers Log Analysis

Run #26606051666 · Attempt 1

日志分析报告

流水线名称	问题标签	修复建议	日志片段
Fleet Model Test / GLM4.5 sft	Blocking Queue 崩溃	DataLoader worker 抛异常导致 blocking queue 被 kill（exit code 241），检查 glm45_sft.sh 数据管道配置，或 rerun 验证是否偶现	报错代码
Fleet Model Test / GLM4.5 lora	检查点文件缺失	glm_full_pp_ckpts 目录下缺少 config.json，sft step 异常退出导致 checkpoint 未写入，lora 依赖该 checkpoint 故连锁失败	报错代码
Fleet Model Test / GLM4.5 dpo	检查点文件缺失	同 lora，run_dpo 尝试 AutoConfig.from_pretrained(/workspace/glm45_fleet/checkpoints/glm_full_pp_ckpts) 找不到 config.json，根因同为 sft 未生成 checkpoint	报错代码

失败的测试 case:

Fleet Model Test - Integration test (H20, multi-card)
  ├── GLM4.5 sft         [FAILED] exit code 241 → SystemError: Blocking queue is killed
  ├── GLM4.5 lora        [FAILED] exit code 1   → FileNotFoundError: config.json not found in glm_full_pp_ckpts
  └── GLM4.5 dpo         [FAILED] exit code 1   → FileNotFoundError: config.json not found in glm_full_pp_ckpts

根本原因分析:

PR #4565（add-glm45-air-128k-yaml）新增了 GLM-4.5-Air 128K benchmark 配置，但本次 CI 是 GLM4.5 基础集成测试。真正的根因是 GLM4.5 sft 训练在第 9 步完成后，DataLoader 的 blocking queue 被 worker 异常 kill（exit code 241），导致 ***** train metrics ***** 未输出、checkpoint /workspace/glm45_fleet/checkpoints/glm_full_pp_ckpts 未写入，后续 lora / dpo 两个 step 均依赖该 checkpoint 做 AutoConfig.from_pretrained，因此连锁报 FileNotFoundError。该问题与 PR 改动（仅加 yaml 配置）直接关联性低，更可能是 Blocking queue is killed 高并发偶现问题。

修复建议:

1. 优先 rerun 当前 CI job，确认 Blocking queue is killed（exit code 241）是否为偶现问题（参考规则：该错误属已知高并发偶现）。
2. 若 rerun 仍复现，排查 glm45_sft.sh 中 DataLoader worker 配置：检查 num_workers、persistent_workers 或数据集读取是否有异常，定位 worker 抛出的原始异常（Blocking queue 错误本身仅是上层捕获，真实错误在 worker 进程）。
3. lora / dpo 失败为 sft 连锁影响，sft 修复后自动恢复，无需单独处理。

---

_{🔄 每次 Re-run 后自动更新}