[Feature] Support MLA dummy load & optimize MLA value padding memory

[chang-wenbin]

Motivation

MLA KV 投影层（MLAKVCombinedProjection）在 dummy 加载模式（load_choices == "dummy"）下，k_b_proj_weight 和 v_b_proj_weight 未在 __init__ 中创建，导致后续推理时访问未定义属性而报错。此外，当 qk_head_dim == v_head_dim 时，对 value 执行 0-padding 是无效操作，浪费内存带宽。

Modifications

• fastdeploy/model_executor/layers/linear.py：在 __init__ 中，当 load_choices == "dummy" 时，创建形状分别为 [num_heads_per_partition, qk_nope_head_dim, kv_lora_rank] 和 [num_heads_per_partition, kv_lora_rank, v_head_dim] 的零初始化 k_b_proj_weight / v_b_proj_weight 参数。
• fastdeploy/model_executor/models/deepseek_v3.py：MLA prefill 路径中，仅当 qk_head_dim != v_head_dim 时执行 value padding，避免零长度 pad 操作。

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.

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

chang-wenbin seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-05-11 17:50:04

📋 Review 摘要

PR 概述：为 MLA KV 投影层补充 dummy 加载模式下的权重初始化，并优化 DeepSeek V3 前向中 value padding 的无效内存操作。
变更范围：fastdeploy/model_executor/layers/linear.py、fastdeploy/model_executor/models/deepseek_v3.py
影响面 Tag：[Models] [OP] [Loader]

---

📝 PR 规范检查

标题缺少官方 Tag，描述各段落均为空（仅含模板占位注释）。

标题建议（可直接复制）：

• [Feature] Support MLA dummy load & optimize MLA value padding memory

PR 描述建议（可直接复制，必须复刻 checklist §D2 模板的完整结构）：

## Motivation
MLA KV 投影层（`MLAKVCombinedProjection`）在 dummy 加载模式（`load_choices == "dummy"`）下，`k_b_proj_weight` 和 `v_b_proj_weight` 未在 `__init__` 中创建，导致后续推理时访问未定义属性而报错。此外，当 `qk_head_dim == v_head_dim` 时，对 value 执行 0-padding 是无效操作，浪费内存带宽。

## Modifications
- `fastdeploy/model_executor/layers/linear.py`：在 `__init__` 中，当 `load_choices == "dummy"` 时，创建形状分别为 `[num_heads_per_partition, qk_nope_head_dim, kv_lora_rank]` 和 `[num_heads_per_partition, kv_lora_rank, v_head_dim]` 的零初始化 `k_b_proj_weight` / `v_b_proj_weight` 参数。
- `fastdeploy/model_executor/models/deepseek_v3.py`：MLA prefill 路径中，仅当 `qk_head_dim != v_head_dim` 时执行 value padding，避免零长度 pad 操作。

## 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
📝 PR 规范	—	标题缺 Tag；Motivation / Modifications / Usage / Accuracy Tests 段落为空
🟡 建议	linear.py	process_weights_after_loading 未对 dummy 模式做早返回，self.kv_b_proj 为 None 时若被调用会抛 AttributeError

process_weights_after_loading 建议修复：

def process_weights_after_loading(self):
    if self.fd_config.load_config.dynamic_load_weight:
        return
    if self.fd_config.load_config.load_choices == "dummy":
        return  # k_b_proj_weight / v_b_proj_weight 已在 __init__ 零初始化，无需再处理
    w = (
        self.kv_b_proj.weight.transpose([1, 0])
        ...
    )

---

总体评价

逻辑改动简洁清晰，dummy 加载初始化与条件 padding 优化均方向正确。建议补充 process_weights_after_loading 的 dummy 守卫以提升健壮性，同时补全 PR 描述。

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-05-11 18:10:34

CI报告基于以下代码生成（30分钟更新一次）:

• PR commit: b81c275
• Merge base: a620e91 (branch: develop)
• 查看完整 Diff
• CI 详情

---

1 任务总览

CI 正在运行中，当前 Required 任务无失败（4 个运行中，1 个等待中）；3 个 Optional 任务失败，不阻塞合并。

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
35(0)	35	26	3	4	2	0

2 任务状态汇总

2.1 Required任务 : 5/10 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
⏳	Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage	-	运行中	-	Job	-
⏳	Extracted partial CE model tasks to run in CI. / run_ce_cases	-	运行中	-	Job	-
⏳	Run Four Cards Tests / run_4_cards_tests	-	运行中	-	Job	-
⏳	xpu_4cards_case_test / run_xpu_4cards_cases	-	运行中	-	Job	-
⏸️	xpu_8cards_case_test / run_xpu_8cards_cases	-	等待中	-	-	-
✅	其余 5 个必选任务通过	-	-	-	-	-

2.2 可选任务 — 21/25 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
❌	Run iluvatar Tests / run_iluvatar_cases	10m29s	Job	-
❌	Check PR Template	11s	Job	-
❌	Trigger Jenkins for PR	21s	Job	-
⏸️	CI_HPU	-	-	-
✅	其余 21 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 0% with 5 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@a620e91). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/layers/linear.py	0.00%	2 Missing and 1 partial ⚠️
fastdeploy/model_executor/models/deepseek_v3.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7775   +/-   ##
==========================================
  Coverage           ?   71.51%           
==========================================
  Files              ?      396           
  Lines              ?    55828           
  Branches           ?     8724           
==========================================
  Hits               ?    39927           
  Misses             ?    13152           
  Partials           ?     2749           

Flag	Coverage Δ
GPU	71.51% <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.