[Feature] Support GQA SWA attention and v_head_dim KV cache

[chang-wenbin]

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.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 66.87117% with 54 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@ecd9733). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/layers/linear.py	54.09%	26 Missing and 2 partials ⚠️
...l_executor/layers/attention/append_attn_backend.py	52.38%	7 Missing and 3 partials ⚠️
fastdeploy/cache_manager/v1/cache_controller.py	76.66%	6 Missing and 1 partial ⚠️
fastdeploy/worker/gpu_model_runner.py	84.84%	4 Missing and 1 partial ⚠️
...eploy/model_executor/layers/attention/attention.py	77.77%	1 Missing and 1 partial ⚠️
fastdeploy/config.py	50.00%	0 Missing and 1 partial ⚠️
...astdeploy/model_executor/ops/triton_ops/do_rope.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8041   +/-   ##
==========================================
  Coverage           ?   67.75%           
==========================================
  Files              ?      475           
  Lines              ?    66694           
  Branches           ?    10284           
==========================================
  Hits               ?    45189           
  Misses             ?    18613           
  Partials           ?     2892           

Flag	Coverage Δ
GPU	77.76% <66.87%> (?)
XPU	6.97% <1.22%> (?)

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 | 2026-06-13 01:00:28

📋 Review 摘要

PR 概述：支持 GQA/SWA attention、per-layer KV head attention backend，以及 v_head_dim 感知的 KV cache / QKV loader 路径。
变更范围：config.py、attention backend/op、KV cache manager、GPU model runner、QKV/QKVG linear loader、PaddleFormers 配置同步及相关测试。
影响面 Tag：[FDConfig] [OP] [KVCache] [Models] [Loader]

问题

级别	文件	概述
🔴 Bug	fastdeploy/model_executor/layers/linear.py:799	shared KV 的非 fused Q/K/V 加载仍按 head_dim 切 V 权重，v_head_dim != head_dim 时会切错或 reshape 失败

历史 Findings 修复情况

Finding	问题	状态
F1	QKVGateParallelLinear 的 TP 分片加载会调用不存在的 _get_kv_shard_id()	⚠️ 仍存在
F2	sliding attention 配置会被当前 window_attn_skip_freq 判断绕过	⚠️ 仍存在
F3	num_key_value_heads_list 未传递到 QKV 投影层	⚠️ 仍存在
F4	QKVGateParallelLinear.qkv_weight_loader() 的 is_scale=True param shard offset/size 未同步缩放	✅ 已修复

📝 PR 规范检查

标题已补官方 [Feature] Tag；PR 描述仍保留模板占位内容，Checklist 未按实际情况勾选。

标题建议（可直接复制）：

• [OP] Support GQA SWA attention and v_head_dim KV cache

PR 描述建议（点击展开，可直接复制）

## Motivation
Support GQA/SWA attention variants, per-layer KV head counts, and `v_head_dim` aware KV cache/weight loading paths.

## Modifications
- Add group size 3 append attention template dispatch coverage.
- Support per-layer attention backends and per-layer KV cache shapes.
- Add `v_head_dim` propagation to config, PaddleFormers sync, attention cache shape, RoPE/cache write path, and QKV/QKVG linear loading.
- Adjust append attention SWA handling and related tests.

## 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.
- [x] 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

本轮按风险优先覆盖了 attention/KV cache/loader 关键路径，并确认最近提交修复了一个历史量化 scale offset 问题。但 v_head_dim 新增语义在非 fused Q/K/V loader 的 shared KV 分支仍有明确错误，需要修复后再合入。