[OPTIMIZATION] save some compute in dsv3

[zhoutianzi666]

Motivation

在 prefill+decode 混合 batch 场景下，MLA decode 路径原本在 kv_b_proj_bmm GEMM 完成后才通过 extract_decoder_token_from_q 筛选 decoder token。本 PR 将筛选提前至 GEMM 之前（在 deepseek_v3.forward 中执行），使昂贵的 GEMM 只作用于 decoder token，节省 encoder token 部分的无效计算。同时清理 mla_blackwell 中遗留的硬编码开发路径，并将 insert_decoder_result_back 的内存分配从 paddle.zeros 改为 paddle.empty 减少初始化开销。

Modifications

• fastdeploy/model_executor/models/deepseek_v3.py：在 need_do_decode 分支（flash_mla/Blackwell 路径），将 extract_decoder_token_from_q 提前至 kv_b_proj_bmm 之前，仅对 decoder token 做 GEMM；GEMM 后调用 insert_decoder_result_back 恢复完整 token 序列。
• fastdeploy/model_executor/layers/attention/mla_attention_backend.py：移除 forward_mixed 内重复的 extract_decoder_token_from_q 调用；insert_decoder_result_back 改用 paddle.empty；删除 mla_blackwell 中硬编码的 /root/... sys.path；Blackwell 判断从 cc >= 100 改为 prop.major == 10；新增 shape 一致性断言。
• tests/operators/test_flashmla_precision.py：新增 SM100/SM90 分支精度测试，引入 page_size 变量替换硬编码的 64。
• tests/operators/test_deepgemm_precision.py：更新 DeepGEMM Blackwell 测试，支持 2CTA 指令和 TMA pipeline。

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.

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-05-13 17:42:32

CI报告基于以下代码生成（30分钟更新一次）:

• PR commit: 2a8d2c6
• Merge base: f01bcde (branch: develop)
• 查看完整 Diff
• CI 详情

---

1 任务总览

所有 Required 任务已通过 ✅，建议合并；有 1 个可选任务失败（不阻塞合并），2 个任务运行中，2 个任务等待中。

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
15(0)	15	10	1	2	2	0

2 任务状态汇总

2.1 Required任务 : 2/2 通过

必选任务阻塞合并，失败需优先处理。

状态	任务	耗时	根因	修复建议	日志	重跑
✅	其余 2 个必选任务通过	-	-	-	-	-

2.2 可选任务 — 8/13 通过

可选任务不阻塞合并，失败仅供参考。

状态	任务	耗时	日志	重跑
❌	Check PR Template	14s	Job	-
⏳	FD-Build-Linux / fd-build	-	Job	-
⏳	Trigger Jenkins for PR	-	Job	-
⏸️	CI_HPU	-	-	-
⏸️	Run iluvatar Tests / run_iluvatar_cases	-	-	-
✅	其余 8 个可选任务通过	-	-	-

---

3 失败详情（仅 required）

无 required 失败任务。

[codecov-commenter]

Codecov Report

❌ Patch coverage is 28.57143% with 20 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@f01bcde). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/models/deepseek_v3.py	38.88%	9 Missing and 2 partials ⚠️
...executor/layers/attention/mla_attention_backend.py	10.00%	7 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7785   +/-   ##
==========================================
  Coverage           ?   63.98%           
==========================================
  Files              ?      461           
  Lines              ?    64145           
  Branches           ?     9826           
==========================================
  Hits               ?    41041           
  Misses             ?    20279           
  Partials           ?     2825           

Flag	Coverage Δ
GPU	73.22% <28.57%> (?)
XPU	7.13% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 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 | 2026-05-13 17:52:26

📋 Review 摘要

PR 概述：将 DSv3 MLA decode 路径中 extract_decoder_token_from_q 提前至 kv_b_proj_bmm GEMM 之前，节省混合 batch 中 encoder token 的无效计算
变更范围：model_executor/models/deepseek_v3.py、layers/attention/mla_attention_backend.py、tests/operators/
影响面 Tag：[Models] [OP] [Optimization]

📝 PR 规范检查

标题使用了 [OPTIMIZATION]（全大写），与官方 Tag [Optimization] 大小写不符；PR 描述完全为空，缺少所有必填 section。

标题建议（可直接复制）：

• [Optimization] save some compute in DSv3 MLA decode path

PR 描述建议（可直接复制）：

## Motivation
在 prefill+decode 混合 batch 场景下，MLA decode 路径原本在 `kv_b_proj_bmm` GEMM 完成后才通过 `extract_decoder_token_from_q` 筛选 decoder token。本 PR 将筛选提前至 GEMM 之前（在 `deepseek_v3.forward` 中执行），使昂贵的 GEMM 只作用于 decoder token，节省 encoder token 部分的无效计算。同时清理 `mla_blackwell` 中遗留的硬编码开发路径，并将 `insert_decoder_result_back` 的内存分配从 `paddle.zeros` 改为 `paddle.empty` 减少初始化开销。

## Modifications
- `fastdeploy/model_executor/models/deepseek_v3.py`：在 `need_do_decode` 分支（flash_mla/Blackwell 路径），将 `extract_decoder_token_from_q` 提前至 `kv_b_proj_bmm` 之前，仅对 decoder token 做 GEMM；GEMM 后调用 `insert_decoder_result_back` 恢复完整 token 序列。
- `fastdeploy/model_executor/layers/attention/mla_attention_backend.py`：移除 `forward_mixed` 内重复的 `extract_decoder_token_from_q` 调用；`insert_decoder_result_back` 改用 `paddle.empty`；删除 `mla_blackwell` 中硬编码的 `/root/...` sys.path；Blackwell 判断从 `cc >= 100` 改为 `prop.major == 10`；新增 shape 一致性断言。
- `tests/operators/test_flashmla_precision.py`：新增 SM100/SM90 分支精度测试，引入 `page_size` 变量替换硬编码的 64。
- `tests/operators/test_deepgemm_precision.py`：更新 DeepGEMM Blackwell 测试，支持 2CTA 指令和 TMA pipeline。

## 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.

问题

级别	文件	概述
🔴 Bug	mla_attention_backend.py:862	SM8x 回归：prop.major == 9 使 A100/L40S（SM89）在 USE_FLASH_MLA=0 时进入 flash_mla 路径并 ImportError
🔴 Bug	test_flashmla_precision.py:93	decoder_res 未定义：非 SM9/SM10 硬件上此行抛 NameError
📝 PR 规范	—	标题 Tag 大小写不符（[OPTIMIZATION] → [Optimization]），描述完全为空

总体评价

优化思路正确，将 GEMM 范围从全 token 收窄到 decoder-only 能有效降低混合 batch 计算量；但 SM 版本条件的收窄（cc >= 100 → prop.major == 9）可能引入 A100（SM80）、L40S（SM89）架构回归，建议修复后合入。

[EmmonsCurse]

LGTM～ Skip coverage check as it mainly relies on tests with sm_version >= 100