[models] support fleet fallback with SWA

[xiaoguoguo626807]

Motivation

Support PaddleFleet fallback models in SWA, attention sink, and EP idle-rank scenarios.

Modifications

• fastdeploy/model_executor/models/paddleformers/base_fleet.py: update PaddleFleet fallback attention patching for SWA, softmax_offset sinks, MLA prefill value padding, and EP idle-rank fake-token forward participation.
• fastdeploy/worker/gpu_model_runner.py: derive per-layer KV head counts from window_attn_skip_freq and swa_num_key_value_heads.
• tests/model_executor/fallback/test_fallback_fleet_model_coverge.py and tests/worker/test_gpu_model_runner.py: add fallback/SWA/sinks/EP idle-rank and per-layer KV head coverage.

Usage or Command

N/A

Accuracy Tests

N/A

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.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 96.82540% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@bff3f75). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...l_executor/layers/attention/append_attn_backend.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8101   +/-   ##
==========================================
  Coverage           ?   67.42%           
==========================================
  Files              ?      474           
  Lines              ?    66946           
  Branches           ?    10302           
==========================================
  Hits               ?    45141           
  Misses             ?    18920           
  Partials           ?     2885           

Flag	Coverage Δ
GPU	77.45% <96.82%> (?)
XPU	6.95% <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.

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-07-16 16:55:11

📋 Review 摘要

PR 概述：支持 PaddleFleet fallback 在 SWA、attention sinks、EP idle-rank 等场景下复用 FastDeploy attention。
变更范围：PaddleFleet fallback attention patch、GPUModelRunner per-layer KV head 初始化、fallback/worker 单测。
影响面 Tag：[Models] [OP] [Executor]

问题

级别	文件	概述
🔴 Bug	fastdeploy/model_executor/models/paddleformers/base_fleet.py:49	FD_FALLBACK_FLEET_USE_ERNIE=0 仍会被当作开启，普通 fallback 可能误走 Ernie 加载路径
🟡 建议	tests/model_executor/fallback/test_fallback_fleet_model_coverge.py:1636	新增 patch_paddlefleet_core_attention 测试 patch 了错误的 Attention 符号

历史 Findings 修复情况

Finding	问题	状态
F1	PaddleFleet backend 默认切到 Ernie 专用加载路径	🔄 部分修复
F2	SWA 层未正确使用 swa_num_key_value_heads	⚠️ 仍存在
F3	混合 KV-head 列表与共享 attention buffer 分配不一致	⚠️ 仍存在
F4	tp_group is None 且 TP>1 时重复初始化 PaddleFleet model parallel	⚠️ 仍存在
F5	zero-size fake token 未同步 forward_meta	⚠️ 仍存在
F6	MLA prefill v padding 后未裁回 v_head_dim	✅ 已修复

📝 PR 规范检查

标题 Tag 仍不符合官方大小写，当前 [models] 应改为 [Models]；PR 描述结构已按模板补齐。

标题建议（可直接复制）：

• [Models] Support PaddleFleet fallback with SWA

总体评价

本轮发现一个新增加载开关解析 bug 和一个新增单测 patch 目标错误；同时多项历史阻塞问题仍未完全修复。建议先修复上述问题，并补强能实际断言 SWA KV-head / sinks wiring 的测试。

[PaddlePaddle-bot]

🔴 Bug os.environ.get() 返回的是字符串，FD_FALLBACK_FLEET_USE_ERNIE=0 / false 这类常见关闭值在 Python 中仍然为真，会意外进入 Ernie 专用加载路径。

这会让显式关闭开关的部署仍走 if USE_ERNIE: 分支，破坏普通 PaddleFormers fallback。建议显式解析布尔值，例如只接受 1/true/yes 为开启：

USE_ERNIE = os.environ.get("FD_FALLBACK_FLEET_USE_ERNIE", "0").lower() in {"1", "true", "yes"}

[PaddlePaddle-bot]

🟡 建议 这里 patch 的是 fastdeploy.model_executor.layers.attention.attention.Attention，但被测函数在 base_fleet.py 模块加载时已经把 Attention 导入并绑定为本地符号，调用时实际使用的是 base_fleet.Attention。

因此这些新增断言拿不到 mock_attention_cls.call_args_list，在安装 paddlefleet 的环境里会失败；即使不失败，也没有验证 patch_paddlefleet_core_attention 真正创建的对象。建议改为 patch fastdeploy.model_executor.models.paddleformers.base_fleet.Attention，并为多层场景使用两个 mock instance 后断言 SWA 层的 kv_num_heads。