[Models] Support MLA_SWA functionality

[chang-wenbin]

Motivation

Support MLA sliding window attention (MLA_SWA) functionality for DeepSeek V3 style MLA layers.

Modifications

• Add SWA top-k indexer construction and a static SWA attention path in fastdeploy/model_executor/models/deepseek_v3.py.
• Add DSAAttentionBackend.forward_static() and reuse it for DSA/MLA SWA FlashMLA calls.
• Adjust MLA KV cache shape/dtype for layers selected by window_attn_skip_freq.
• Update GPU model runner KV cache allocation to create uint8 MLA cache for SWA layers.

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 26.92308% with 95 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@cbb0811). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/models/deepseek_v3.py	8.04%	80 Missing ⚠️
fastdeploy/worker/gpu_model_runner.py	45.45%	4 Missing and 2 partials ⚠️
...executor/layers/attention/dsa_attention_backend.py	80.76%	2 Missing and 3 partials ⚠️
...executor/layers/attention/mla_attention_backend.py	33.33%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8049   +/-   ##
==========================================
  Coverage           ?   67.47%           
==========================================
  Files              ?      475           
  Lines              ?    66739           
  Branches           ?    10293           
==========================================
  Hits               ?    45031           
  Misses             ?    18831           
  Partials           ?     2877           

Flag	Coverage Δ
GPU	77.43% <26.92%> (?)
XPU	6.97% <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 | 2026-06-16 02:13:02

总体评价

本轮按风险优先审查了 MLA_SWA 的 attention 前向、SWA indexer、KV cache 分配和新增测试。当前没有新增阻断级 inline finding，但这次改动覆盖 DeepSeekV3 MLA/DSA 的核心前向路径和 KV cache layout，PR 仍缺少能证明 MLA_SWA 语义正确性的精度/回归测试；建议合入前补齐。

问题

• 🟡 建议：请补充 MLA_SWA 的真实语义/精度回归覆盖。当前测试主要验证 mock 后的 DSA 输出 shape 和默认 MLA cache shape，没有覆盖 window_attn_skip_freq 命中层时的 forward_swa_static()、SWA top-k indexer、prefill/decode/mixed batch 输出，尤其没有与非 SWA/参考实现对齐的 logits 或 attention 输出证据。这个 PR 改了模型核心 attention 与 KV cache dtype/shape，建议至少补充 SWA layer 的 prefill、decode、mixed batch 用例，以及 sliding_window 边界和 prefix/chunked prefill 场景的对齐结果。

PR 规范检查

标题 [Models] Support MLA_SWA functionality 符合 Tag 规范，PR 描述结构完整。Accuracy Tests 目前为 N/A，但本 PR 属于模型前向语义变更，建议在描述中补充实际精度/对齐结果或明确可复现的验证命令。

历史 Finding 状态

已复查历史未解决项：f_d5017472、f_7eff4a88、f_28e9c24f、f_f824c9a1、f_2ac6ab5a 在当前 diff/代码中仍未完全修复；f_c57369dc 对应的非 SWA 层 KV cache 量化配置被清空问题已通过在 else 分支内恢复 quant_config.kv_cache_quant_type 修复。本轮不重复发送这些历史问题的 inline comment。

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-06-16 10:55:13 UTC+08:00

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: cc20efd | Merge base: cbb0811 (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：diff覆盖率31%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: 覆盖率检查

用例	错误摘要
diff-cover python_coverage_all.xml --fail-under=80	PR diff 覆盖率 31%，低于 80% 阈值

关键日志:

TEST_EXIT_CODE: 0
COVERAGE_EXIT_CODE: 9
Failure. Coverage is below 80%.
fastdeploy/model_executor/layers/attention/mla_attention_backend.py (50.0%): Missing lines 546,626-627
fastdeploy/model_executor/models/deepseek_v3.py (8.0459%): Missing lines 92,94,96-117,128-134,302-303,449-677 等
fastdeploy/worker/gpu_model_runner.py (63.6363%): Missing lines 1656-1657,1668-1669
fastdeploy/model_executor/layers/attention/dsa_attention_backend.py (92.3077%): Missing lines 386,411
total_num_lines: 130, total_num_violations: 89, total_percent_covered: 31
Process completed with exit code 9.

• 根因摘要: 新增 MLA_SWA 代码覆盖率不足
  测试阶段已通过，失败发生在 Verify Code Coverage Threshold (80%)。本 PR 新增/修改的 MLA_SWA 路径、KV cache shape/dtype 分支和 DSA 静态 attention 分支缺少对应测试覆盖，导致 diff-cover 统计 130 个变更行中 89 行未覆盖，总 diff 覆盖率仅 31%。

修复建议:

1. 补充针对 fastdeploy/model_executor/models/deepseek_v3.py 的 MLA_SWA 单测，覆盖 get_swa_indexer_top_k 的 encoder/decoder 分支和 forward_swa_static 调用 DSAAttentionBackend.forward_static 的路径。
2. 扩展 tests/layers/test_mla_attention_kv_cache.py 或新增用例，覆盖 MLAAttentionBackend.get_kv_cache_shape() 在 window_attn_skip_freq[layer_id] == 1 时的 fp8 cache shape 分支。
3. 扩展 GPU runner 初始化相关测试，覆盖 fastdeploy/worker/gpu_model_runner.py 中 MLA SWA 层选择 uint8 cache 的分支；同时补齐 DSAAttentionBackend.forward_static 的 q head padding prefill/decode 分支。

关联变更: fastdeploy/model_executor/models/deepseek_v3.py、fastdeploy/model_executor/layers/attention/mla_attention_backend.py、fastdeploy/model_executor/layers/attention/dsa_attention_backend.py、fastdeploy/worker/gpu_model_runner.py