[Models] support swa for fleet fallback model

[xiaoguoguo626807]

Motivation

Support sliding-window attention (SWA) for PaddleFleet fallback MLA models, so SWA layers selected by window_attn_skip_freq can use the DSA attention path.

Modifications

• fastdeploy/model_executor/models/paddleformers/base_fleet.py: pass window_attn_skip_freq and sliding_window into FastDeployAttention.
• fastdeploy/model_executor/models/paddleformers/base_fleet.py: add an SWA branch that builds indexer_top_k and calls DSAAttentionBackend.forward_static for MLA attention.
• fastdeploy/model_executor/models/paddleformers/base_fleet.py: adjust PaddleFleet fallback config/model loading and MoE-related config fields.

Usage or Command

N/A

Accuracy Tests

N/A（当前 PR 未提供 SWA/fallback logits 对齐或精度测试结果）

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

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@f4eda5a). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8090   +/-   ##
==========================================
  Coverage           ?   67.52%           
==========================================
  Files              ?      475           
  Lines              ?    66958           
  Branches           ?    10329           
==========================================
  Hits               ?    45213           
  Misses             ?    18874           
  Partials           ?     2871           

Flag	Coverage Δ
GPU	77.55% <100.00%> (?)
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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-07-06 14:33:48

📋 Review 摘要

PR 概述：为 PaddleFleet fallback MLA 模型新增基于 window_attn_skip_freq 的 SWA/DSA attention 路径，并调整 fallback 初始化/测试。
变更范围：base_fleet.py、fallback 单测
影响面 Tag：[Models]

问题

级别	文件	概述
🟡 建议	fastdeploy/model_executor/models/paddleformers/base_fleet.py:178	新增 SWA 前向路径缺少 logits/精度对齐证据
🟡 建议	tests/model_executor/fallback/test_fallback_fleet_model_coverge.py:1365	SWA 单测 mock 掉真实 DSA 且未校验真实入参 rank/shape

📝 PR 规范检查

标题包含官方 [Models] Tag，描述包含 checklist.md §D2 要求的 Motivation、Modifications、Usage or Command、Accuracy Tests、Checklist 结构，符合规范。

总体评价

实现方向和 DeepSeek SWA/DSA 路径基本对齐，但当前验证主要停留在控制流/shape mock。建议补齐最小数值对齐和真实后端入参约束后再合入核心模型路径。

[PaddlePaddle-bot]

🟡 建议 这里新增了 window_attn_skip_freq 命中的 MLA/SWA 层直接走 DSAAttentionBackend.forward_static 的前向路径，但 PR 只提供了 mock/shape 级单测，Accuracy Tests 也标为 N/A。该分支会改变 prefill/decode 的 mask、indexer、cache 写入和 softmax scale 使用，输出直接影响 logits；没有 fallback logits 或参考实现对齐，后续 get_swa_indexer_top_k、cache layout、V de-absorb 任一错误都不会被发现。

建议修复方式：
补一个小模型/固定 seed 的 PaddleFleet fallback SWA 对齐用例，至少覆盖 window_attn_skip_freq[layer_id] == 1 的 prefill-only 和 mixed prefill+decode，并与 DeepSeek/PaddleFleet 参考路径比较 logits 容差；如果 CI 无法跑真实 GPU kernel，请在 PR 中附可复现命令和对齐结果。

[PaddlePaddle-bot]

🟡 建议 这个用例把 DSAAttentionBackend.forward_static 整体替换成 MagicMock，同时构造的 kv_compressed squeeze 后是 3D [seq, heads, rank]，但真实 DSA backend 要求 compressed_kv 为 2D、k_pe 为 3D。因此测试会在真实调用会触发 shape/rank 问题时仍然通过。

建议修复方式：
让用例使用真实后端约束的入参形状，例如 kv_compressed squeeze 后为 [seq, kv_lora_rank]、k_pos_emb squeeze 后为 [seq, 1, qk_rope_head_dim]、q_absorbed 最后一维为 kv_lora_rank + qk_rope_head_dim，并对 mock_dsa.forward_static.call_args 断言 q.ndim == 3、compressed_kv.ndim == 2、k_pe.ndim == 3。

[paddle-bot]

Thanks for your contribution!