[Models] Update SWA RoPE theta for MLA/GQA attention

[chang-wenbin]

Motivation

为配置了 swa_rope_theta 的 MLA/GQA 滑动窗口注意力层使用独立 RoPE base，避免 SWA 层与全量注意力层共用 rope_theta。

Modifications

• ForwardMeta 增加 swa_rotary_embs，gpu_model_runner 从 share_inputs["swa_rope_emb"] 传入。
• InputBatch 和 ProposerInputBatch 在配置 swa_rope_theta 时额外构建 swa_rope_emb。
• AppendAttentionBackend 在 window_attn_skip_freq[layer_id] == 1 且配置 swa_rope_theta 时使用 swa_rotary_embs。
• DeepseekV3MLAAttention 对 SWA 层使用 swa_rope_theta 初始化 RoPE，并缓存 window_attn_skip_freq。

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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-25 19:30:12

📋 Review 摘要

PR 概述：为 MLA/GQA 的 SWA 注意力路径接入独立 swa_rope_theta / swa_rope_emb。
变更范围：ForwardMeta、InputBatch / ProposerInputBatch、GPU runner、append attention backend、DeepSeek V3 MLA attention。
影响面 Tag：[Models] [OP]

问题

未发现新的阻塞性问题。PR 规范问题在下面章节报，不要在这里重复。

历史 Findings 修复情况

Finding	问题	状态
F1	glm_moe_dsa 的非 SWA 层会丢失 rope_parameters["rope_theta"]。	⚠️ 仍存在
F2	rope_already_applied 的 PaddleFormers fallback 在 SWA 层会再次应用 RoPE。	⚠️ 仍存在

📝 PR 规范检查

标题缺少官方 Tag，描述仍是空模板且未提供精度/对齐结果；下面给出可直接替换的标题和描述。

标题建议（可直接复制）：

• [Models] Update SWA RoPE theta for MLA/GQA attention

PR 描述建议（点击展开，可直接复制）

## Motivation
为配置了 `swa_rope_theta` 的 MLA/GQA 滑动窗口注意力层使用独立 RoPE base，避免 SWA 层与全量注意力层共用 `rope_theta`。

## Modifications
- `ForwardMeta` 增加 `swa_rotary_embs`，`gpu_model_runner` 从 `share_inputs["swa_rope_emb"]` 传入。
- `InputBatch` 和 `ProposerInputBatch` 在配置 `swa_rope_theta` 时额外构建 `swa_rope_emb`。
- `AppendAttentionBackend` 在 `window_attn_skip_freq[layer_id] == 1` 且配置 `swa_rope_theta` 时使用 `swa_rotary_embs`。
- `DeepseekV3MLAAttention` 对 SWA 层使用 `swa_rope_theta` 初始化 RoPE，并缓存 `window_attn_skip_freq`。

## 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.
- [ ] 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

本轮按风险优先审查了 5 个变更文件中的 RoPE/SWA 主链路，并核对了 append attention backend、DeepSeek MLA、GPU ForwardMeta 初始化和输入 batch 构造。除历史未解决的 F1/F2 外，未发现新的 diff 级别阻塞缺陷；合入前仍建议修复这两个历史问题并补充精度/对齐说明。

[codecov-commenter]

Codecov Report

❌ Patch coverage is 31.81818% with 15 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@e8ae0f9). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/input_batch.py	33.33%	4 Missing and 4 partials ⚠️
fastdeploy/model_executor/models/deepseek_v3.py	0.00%	5 Missing ⚠️
...l_executor/layers/attention/append_attn_backend.py	50.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8077   +/-   ##
==========================================
  Coverage           ?   67.62%           
==========================================
  Files              ?      475           
  Lines              ?    66909           
  Branches           ?    10321           
==========================================
  Hits               ?    45249           
  Misses             ?    18813           
  Partials           ?     2847           

Flag	Coverage Δ
GPU	77.53% <31.81%> (?)
XPU	16.03% <4.54%> (?)

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 | 2026-06-26 11:48:07 UTC+08:00

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: a85a97a | Merge base: e8ae0f9 (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
68(26)	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：差异覆盖率 50% 未达 80%	高	Job

2 失败详情

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

分析器: 通用分析(fallback)

失败用例: 覆盖率阈值检查

用例	错误摘要
Verify Code Coverage Threshold (80%)	新增代码差异覆盖率为 50%，低于 80% 阈值，进程以 exit code 9 失败

关键日志:

GPU Patch Coverage Details:
"fastdeploy/model_executor/models/deepseek_v3.py": percent_covered=0.0, violation_lines=[302, 304, 311, 316, 536]
"fastdeploy/worker/input_batch.py": percent_covered=66.67, violation_lines=[252, 742, 854, 1092]
"fastdeploy/model_executor/layers/attention/append_attn_backend.py": percent_covered=50.0, violation_lines=[327, 328]
"total_num_lines": 22, "total_num_violations": 11, "total_percent_covered": 50
Process completed with exit code 9.

• 根因摘要: 新增 SWA RoPE 逻辑缺少覆盖率
  PR 新增了 swa_rope_theta、swa_rope_emb 和 SWA 层切换 swa_rotary_embs 的逻辑，但对应分支没有被当前单测覆盖。单测执行本身通过，失败发生在覆盖率门禁，说明需要补充覆盖这些新增分支的测试或调整已有测试触发相关配置。

修复建议:

1. 为 deepseek_v3.py 中 swa_rope_theta 覆盖 self.rope_theta 以及 SWA 分支判断补测试，重点覆盖第 302、304、311、316、536 行。
2. 为 input_batch.py 中 swa_rope_emb = get_rope(...) 的初始化和 reset 路径补测试，覆盖第 252、742、854、1092 行。
3. 为 append_attn_backend.py 中 SWA 层选择 forward_meta.swa_rotary_embs 的分支补测试，覆盖第 327、328 行。

关联变更: fastdeploy/model_executor/models/deepseek_v3.py, fastdeploy/worker/input_batch.py, fastdeploy/model_executor/layers/attention/append_attn_backend.py