[codex] Adapt DSv4 backward alignment to latest PaddleFormers

[SigureMo]

Summary

Port the DSv4 backward-alignment changes onto the latest PaddleFormers develop baseline.

This branch includes the two backward-alignment commits ported from the old alignment branch, plus local run-entry/data-path adaptation used by the latest 10-step alignment run:

• keep the SFT precision run from saving final model/state artifacts
• add DSV4_FLEET_FIXED_TOKENS JSON fixed-token batch support in collate_fn
• preserve the existing LOAD_FIXED_DATA_PATH path and print which fixed-data source is used

This PR is intended to be used together with the matching PaddleFleet PR from branch codex/dsv4-backward-align-latest.

Validation

• git diff --check origin/develop..HEAD
• 10-step DSv4 old/latest comparison completed with paired PaddleFleet changes:
  • main loss md5 diff: 0/10
  • MTP loss md5 diff: 0/10
  • latest log: /root/paddlejob/share-storage/gpfs/system-public/huangjiyi/dsv4_backward_align/0605_latest_align_22steps/outputs/dpskv4_ep8_4layer_1k_align_20260606_1318/output_0/workerlog.0
  • old log: /root/paddlejob/share-storage/gpfs/system-public/huangjiyi/dsv4_backward_align/0605_align_22steps/outputs/dpskv4_ep8_4layer_1k_align_20260606_1325/output_0/workerlog.0

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

root seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Paddle-CI-Bot]

PaddleFormers Log Analysis

Run #27056714883 · 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 - Integration test (H20, single card)	PP段切分 Bug	glm4_moe/modeling.py 中 Glm4MoeDecoderLayer 层名与 paddlefleet _gen_layer_weight 正则匹配失败，检查 gpt_provider.py 传入的 seg_method 或 decoder 层类名是否与 paddlefleet 期望一致	报错代码
Fleet Model Test - Integration test (H20, multi-card)	PP段切分 Bug	同上；另有 EP4 场景出现 number of layers (1) should be divided by part number(2)，说明 GLM4.5 MoE DSA 的层数配置与 PP stage 数不整除，需检查 num_layers 或 PP size 设置	报错代码
Fleet Model Test - Integration test (A100)	PP段切分 Bug	与 H20 同根因，GLM4.5 所有 task（pre-train/sft/lora/dpo/dpo_lora）均因 weight_idxs 为空或层数不整除失败	报错代码
Unittest GPU CI - unittest-gpu-ci	PP段切分 Bug + SwanLab 初始化缺失	1) GlmMoeDsaModelTest 同 Fleet 根因；2) SwanLabCallback.on_train_begin 未先调用 swanlab.init()，测试需在 mock swanlab.get_run() 或在 setUp 中先 init	报错代码

失败的测试 case:

# Fleet Model Test (H20 single-card)
- Integration test (GLM4.5 single-card)        → AssertionError: weight_idxs' length should be greater than 0
- Integration test (Qwen3-30B-A3B single-card) → AssertionError: weight_idxs' length should be greater than 0

# Fleet Model Test (H20 multi-card)
- GLM4.5 pre-train       → weight_idxs length 0 / exit code 241
- GLM4.5 sft             → exit code 241
- GLM4.5 sft cp          → exit code 1
- GLM4.5 lora            → exit code 1
- GLM4.5 dpo             → exit code 241
- GLM4.5 dpo_lora        → exit code 1
- GLM4.5 pre-train (EP4) → number of layers (1) should be divided by part number(2)
- GLM4.5 pre-train (FP8) → weight_idxs length 0
- GLM4.5 pre-train (Grouped GEMM) → weight_idxs length 0
- Qwen pre-train         → weight_idxs length 0
- Qwen sft               → exit code 1
- Qwen lora              → exit code 1
- Qwen vl lora           → exit code 1

# Fleet Model Test (A100)
- GLM4.5 pre-train       → weight_idxs length 0 / exit code 241
- GLM4.5 sft             → exit code 241
- GLM4.5 lora            → exit code 1
- GLM4.5 dpo             → exit code 241
- GLM4.5 dpo_lora        → number of layers (1) / exit code 241
- Qwen pre-train         → weight_idxs length 0
- Qwen sft               → exit code 1
- Qwen lora              → exit code 1
- Qwen vl lora           → exit code 1

# Unittest GPU CI
- tests/trainer/test_trainer_visualization.py::TestSwanlabCallback::test_swanlabcallback
- tests/transformers/glm_moe_dsa/test_modeling.py::GlmMoeDsaModelTest::test_GlmMoeDsa_lm_head_model

根本原因分析:

本 PR 引入了 glm_moe_dsa（GLM4.5 MoE DSA）新模型，其 decoder 层命名为 Glm4MoeDecoderLayer，但 gpt_provider.py 传给 gpt_builder 的 seg_method（PP pipeline 分段依据的层名 pattern）仍然指向旧的层名（如 Glm4MoeDecoderLayer 无法被 paddlefleet _gen_layer_weight 中 regex.search(name) 匹配），导致 weight_idxs 为空，所有使用该 GPTProvider 路径的 GLM4.5 MoE 场景（pre-train / sft / lora / dpo 以及 Unittest 中的 GlmMoeDsaModelTest）全部崩溃；SwanLabCallback 测试则是新 case 忘记在 setUp 中初始化 swanlab run。

修复建议:

1. PP 层名匹配修复：在 paddleformers/transformers/gpt_provider.py 的 provide() 方法中，确认 seg_method 传入的 layername 与 GlmMoeDsaForCausalLM 实际使用的 decoder 层类名一致。排查方式：在 glm_moe_dsa/modeling.py 中打印 self._layers_desc 各层 __name__，确认 regex pattern 能命中。若 DSA 版本将 decoder 层改名（如改为 GlmMoeDsaDecoderLayer），则需同步更新 seg_method 中的 pattern 字符串。

2. EP4 层数整除问题：number of layers (1) should be divided by part number(2) 说明 CI 配置里 num_layers 设置为 1 而 PP size=2，需要检查 tests/config/ci/glm45_*_ep4*.yaml 中 num_hidden_layers 是否被正确配置为可被 PP stage 数整除的值。

3. SwanLabCallback 测试修复：在 tests/trainer/test_trainer_visualization.py 的 TestSwanlabCallback.setUp 中添加 swanlab.init(mode="disabled") 或 mock swanlab.get_run() 返回一个假 run 对象，避免 on_train_begin 中因无 active run 抛出 RuntimeError。

---

🔍 准确性记录：请点击评论底部 😊 图标，选择 👍（准确）或 👎（有误），将自动记录到 CI 监控系统

_{🔄 每次 Re-run 后自动更新}