[Optimization][DeepSeekV3.2]Reducing slot_mapping compute frequency from twice per layer to a single pre-processing step.

[ShaneGZhu]

Motivation

使用5层DeepSeekV3.2-Exp-BF16进行Profile
--------------关闭CudaGraph之后的性能提升--------------------
V0版本：图中decode每个Step耗时17.2ms，每层layer耗时3.2ms，其中slotmapping在DSA_Index和DSA_attn中占据不少时间：

• 一个layer中两次耗时加起来大约300us，占比0.3ms/3.282ms=9.1%
• 在一个Step中，五层要耗时1.5ms，占比 1.5/17.2 = 8.7%
  

注意：最终目标是将preprocess逻辑放到了gpu_model_runner中，后续计划作为每一个Step通用的逻辑，而不仅仅是某一个模型的Trick。

V1版本 ：关闭CudaGraph +消除冗余计算

看到DSA_Indexer和DSA_attn已经没有了slot计算耗时。
结果：图中decode每个Step耗时降低到15.8ms，相比V0版本降低1.4ms（+8.13%）

--------------保持CudaGraph开启之后的性能提升--------------------
V2版本 开启CudaGraph +slotmapping冗余计算：

之后图中decode每个Step耗时约7.2ms
V3版本 开启CudaGraph +消除冗余计算：

图中decode每个Step耗时约6.2ms，相比V2版本降低1ms（+13.88%）

💡 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

本次优化将 DeepSeekV3.2 模型中 slot_mapping 的计算从每层两次减少为单次预处理，显著减少重复计算，并将通用的slotmapping和postionids移到了gpu_model_runner层。
核心变更修改文件列表 (5个文件):

• fastdeploy/model_executor/forward_meta.py
• fastdeploy/model_executor/layers/attention/dsa_attention_backend.py
• fastdeploy/model_executor/models/deepseek_v3.py
• fastdeploy/worker/gpu_model_runner.py
• fastdeploy/worker/input_batch.py
  次要修改：删除了get_position_ids_and_mask_encoder_batch算子不必要的参数，调整单测文件

Usage or Command

None

Accuracy Tests

评测集	指标	Before	After
IFEval_gen	Prompt-level-strict-accuracy	86.69	87.8
	Inst-level-strict-accuracy	91.01	91.61
	Prompt-level-loose-accuracy	89.46	90.94
	Inst-level-loose-accuracy	93.05	93.65

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!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 96.00000% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@14d5566). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/models/deepseek_v3.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7367   +/-   ##
==========================================
  Coverage           ?   73.89%           
==========================================
  Files              ?      398           
  Lines              ?    54943           
  Branches           ?     8606           
==========================================
  Hits               ?    40598           
  Misses             ?    11632           
  Partials           ?     2713           

Flag	Coverage Δ
GPU	73.89% <96.00%> (?)

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.

[chang-wenbin]

单测需要check修复下

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-16 15:40 CST

📋 Review 摘要

PR 概述：将 DeepSeekV3.2 模型中 slot_mapping 的计算从每层两次冗余计算优化为在 gpu_model_runner 层的单次预处理，同时清理 mask_encoder_batch 不再使用的参数。
变更范围：gpu_model_runner、input_batch、deepseek_v3 模型、dsa_attention_backend、CUDA 算子
影响面 Tag：Optimization Models OP Executor

问题

级别	文件	概述
🟡 建议	gpu_model_runner.py:1284	attn_backends[0] 缺少空列表防御检查
🟡 建议	gpu_model_runner.py:1296	assert 缺少描述性错误信息

总体评价

优化方案设计合理，将 slot_mapping 和 position_ids 的计算从模型层（deepseek_v3.py）和算子层（dsa_attention_backend.py）提升到 gpu_model_runner 统一预处理，消除了每层重复计算，Profile 数据也验证了性能收益。重复的 compute_slot_mapping 函数被正确移除，mask_encoder_batch 参数的清理也干净彻底。两条小建议均为代码健壮性改进，不阻塞合入。

[PaddlePaddle-bot]

🟡 建议 此处直接访问 self.attn_backends[0] 未检查列表是否为空。

对比同文件 _process_reorder 方法中的写法 if self.attn_backends and getattr(...) 做了空列表保护。虽然 _initialize_attn_backend 在 __init__ 中先于此方法被调用，保证了 attn_backends 非空，但建议保持一致的防御式编程风格。

建议修改：

if not self.attn_backends or not isinstance(self.attn_backends[0], (MLAAttentionBackend, DSAAttentionBackend)):
    return

[PaddlePaddle-bot]

🟡 建议 此处 assert 语句在生产环境中可能被优化掉（python -O 模式），建议添加描述性的错误信息方便排查问题。

assert self.forward_meta.batch_id_per_token.shape == block_idx.shape, (
    f"Shape mismatch: batch_id_per_token {self.forward_meta.batch_id_per_token.shape} "
    f"vs block_idx {block_idx.shape}"
)