[Models] fix fleet model fallback ep init

[xiaoguoguo626807]

Motivation

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-11 15:17:38

📋 Review 摘要

PR 概述：调整 PaddleFleet fallback 的生成校验与 EP/TP 初始化流程。
变更范围：fastdeploy/config.py、fastdeploy/model_executor/models/paddleformers/base_fleet.py
影响面 Tag：[FDConfig] [Models]

问题

级别	文件	概述
🔴 Bug	fastdeploy/model_executor/models/paddleformers/base_fleet.py:420	重新 fleet.init() 后仍可能复用旧的 PaddleFleet TP group

📝 PR 规范检查

标题包含官方 [Models] Tag，符合标题规范；PR 描述的 Motivation、Modifications、Usage or Command、Accuracy Tests 仍是模板占位，建议替换为以下完整内容。

标题建议（可直接复制）：

• [Models] Fix PaddleFleet fallback EP initialization

PR 描述建议（点击展开，可直接复制）

## Motivation
修复 `--model-impl paddlefleet` fallback 场景下生成模型校验与 Expert Parallel 初始化状态复用问题，避免 PaddleFleet 初始化时使用不匹配的 TP/EP parallel state。

## Modifications
- `fastdeploy/config.py`：允许 `model_impl=paddlefleet` 在 `runner_type=generate` 且模型未被 FastDeploy 原生 registry 标记为生成模型时继续走 fallback。
- `fastdeploy/model_executor/models/paddleformers/base_fleet.py`：为 PaddleFleet fallback 配置 CPU 初始化和跳过参数初始化，并在 `fleet.init()` 前重置 Paddle Fleet hybrid topology。

## Usage or Command
N/A

## Accuracy Tests
N/A

## Checklist

- [x] Add at least a tag in the PR title.
  - Tag list: [`[FDConfig]`,`[APIServer]`,`[Engine]`, `[Scheduler]`, `[PD Disaggregation]`, `[Executor]`, `[Graph Optimization]`, `[Speculative Decoding]`, `[RL]`, `[Models]`, `[Quantization]`, `[Loader]`, `[OP]`, `[KVCache]`, `[DataProcessor]`, `[BugFix]`, `[Docs]`, `[CI]`, `[Optimization]`, `[Feature]`, `[Benchmark]`, `[Others]`, `[XPU]`, `[HPU]`, `[GCU]`, `[DCU]`, `[Iluvatar]`, `[Metax]`]
  - You can add new tags based on the PR content, but the semantics must be clear.
- [ ] Format your code, run `pre-commit` before commit.
- [ ] Add unit tests. Please write the reason in this PR if no unit tests.
- [ ] Provide accuracy results.
- [x] If the current PR is submitting to the `release` branch, make sure the PR has been submitted to the `develop` branch, then cherry-pick it to the `release` branch with the `[Cherry-Pick]` PR tag.

总体评价

这个 PR 的方向和问题域是合理的，但当前实现会在已有 PaddleFleet TP group 非空时跳过重新初始化，仍可能保留旧拓扑状态。建议先修复该初始化一致性问题，并同步补充/更新 PaddleFleet fallback 初始化路径的回归验证。

[PaddlePaddle-bot]

🔴 Bug 这里只在 _TENSOR_MODEL_PARALLEL_GROUP is None 时才用新的 HCG 初始化 PaddleFleet TP group，会把已有但拓扑已过期的 TP group 保留下来。

本函数前面只重置了 paddle.distributed.fleet.base.topology._HYBRID_PARALLEL_GROUP 和 paddle.distributed.parallel_helper.__parallel_ctx__clz__，没有清理 paddlefleet.parallel_state._TENSOR_MODEL_PARALLEL_GROUP。如果该全局已经由旧的 FastDeploy/Fleet 拓扑初始化过，fleet.init() 重建 HCG 后这里会直接跳过 parallel_state.initialize_model_parallel(hcg)，PaddleFleet 的 Column/RowParallelLinear 仍按旧 TP group 分片，和新的 EP/TP 拓扑不一致。

建议修复方式：在重新 fleet.init() 前同步清理 PaddleFleet parallel_state 中的 TP group/global ranks，或保留旧代码里的 group size/topology mismatch 检查；发现当前 group 与 parallel_config.tensor_parallel_size 或新 HCG 不一致时，必须用 fleet.get_hybrid_communicate_group() 重新调用 parallel_state.initialize_model_parallel(hcg)。

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@f161fea). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8039   +/-   ##
==========================================
  Coverage           ?   67.71%           
==========================================
  Files              ?      471           
  Lines              ?    66341           
  Branches           ?    10211           
==========================================
  Hits               ?    44925           
  Misses             ?    18549           
  Partials           ?     2867           

Flag	Coverage Δ
GPU	77.78% <100.00%> (?)
XPU	6.99% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-06-12 21:32:45

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: f5a72c4 | Merge base: f161fea (branch: develop)

---

1 Required任务 : 9/10 通过

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
42(0)	42	38	4	0	0	0

任务	错误类型	置信度	日志
Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage	PR问题	高	Job

2 失败详情

🔴 Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage — PR问题（置信度: 高）

错误类型: PR问题 | 置信度: 高
分析器: 通用分析(fallback)
失败用例:

用例	错误摘要
tests/model_executor/fallback/test_fallback_fleet_model_coverge.py::TestInitPaddlefleetParallelState::test_seed_assertion_error_is_silenced	initialize_model_parallel 触发 PaddleFleet 全局 memory buffer 已初始化断言

关键日志:

tests/model_executor/fallback/test_fallback_fleet_model_coverge.py:1070: model._init_paddlefleet_parallel_state(fd_config)
fastdeploy/model_executor/models/paddleformers/base_fleet.py:422: parallel_state.initialize_model_parallel(hcg)
/usr/local/lib/python3.10/dist-packages/paddlefleet/parallel_state.py:426: assert _GLOBAL_MEMORY_BUFFER is None
AssertionError: global memory buffer is already initialized

• 根因摘要: PR改为初始化PaddleFleet时未重置全局缓冲
  PR 在 fastdeploy/model_executor/models/paddleformers/base_fleet.py:420-422 删除了原先 TP=1 手动建组/TP size mismatch 分支，改为只要 _TENSOR_MODEL_PARALLEL_GROUP is None 就调用 parallel_state.initialize_model_parallel(hcg)。失败用例将 _TENSOR_MODEL_PARALLEL_GROUP 置空后，本应验证 model_parallel_cuda_manual_seed 的 AssertionError 会被 base_fleet.py:428-431 吞掉，但新逻辑先进入 initialize_model_parallel，被 PaddleFleet 已存在的 _GLOBAL_MEMORY_BUFFER 断言拦截，seed 分支没有执行。因此这是本 PR 的 base_fleet.py 初始化逻辑变更直接触发的单测失败。

修复建议:

1. 在 base_fleet.py 调用 parallel_state.initialize_model_parallel(hcg) 前，使用 PaddleFleet/Paddle 的正式 reset/destroy API 清理已有 parallel_state 全局状态，至少要覆盖 _GLOBAL_MEMORY_BUFFER；若 TP=1 不需要重新初始化 model parallel，则保留或恢复原 TP=1 手动建组路径。
2. 若新初始化语义是预期行为，同步更新 tests/model_executor/fallback/test_fallback_fleet_model_coverge.py:1049 附近的用例，在只验证 seed 断言吞掉时 mock 或重置 ps.initialize_model_parallel 相关全局状态，并清理仍描述旧 TP=1 分支的断言。

关联变更: fastdeploy/model_executor/models/paddleformers/base_fleet.py:399-422；PR 删除旧 TP=1/mismatch 分支并新增 fleet 状态重置后统一初始化。