[BugFix]Fix DSA multi-batch inference deployment

[chang-wenbin]

Motivation

Fix DSA multi-batch inference deployment

💡 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

Modifications

1. indexer_topk.cuh:

   • decode 模式下修改 batch_id 计算方式，从 batch_id_per_token[bid / q_num_heads] 改为直接使用 bid / q_num_heads
   • 修改 dst 指针计算，从 output + bid * top_k 改为 output + aux_input[batch_id] * top_k 以支持 multi-batch 场景

2. deepseek_v3.py:

   • prefill 路径：删除冗余的 logits 重排逻辑，直接使用原始 logits
   • decode 路径：使用 cu_seqlens_q 作为 offsets，cache_seqlens 作为 seq_len_decoder

3. test_radix_topk_accuracy.py:

   • 更新 decode 模式测试用例，使用 cu_seqlens_q 替代 offsets

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!

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-08 18:02 CST

📋 Review 摘要

PR 概述：修复 DSA（DeepSeek Attention）multi-batch decode 推理部署问题，主要修改了 kernel 的 batch_id 计算和输出指针映射逻辑。

变更范围：custom_ops/gpu_ops/sparse_indexer/、fastdeploy/model_executor/models/deepseek_v3.py、tests/operators/

影响面 Tag：[OP] [Models]

📝 PR 规范检查

PR 标题：包含有效的 [BugFix] 标签 ✅

PR 描述：

• Motivation 部分填写了目的 ✅
• Modifications 部分未填写内容 ⚠️
• Checklist 中的项目均未勾选 ⚠️

建议补充：

描述模板（可直接复制）：

## Modifications

1. `indexer_topk.cuh`: 
   - decode 模式下修改 batch_id 计算方式，从 `batch_id_per_token[bid / q_num_heads]` 改为直接使用 `bid / q_num_heads`
   - 修改 dst 指针计算，从 `output + bid * top_k` 改为 `output + aux_input[batch_id] * top_k` 以支持 multi-batch 场景

2. `deepseek_v3.py`:
   - prefill 路径：删除冗余的 logits 重排逻辑，直接使用原始 logits
   - decode 路径：使用 `cu_seqlens_q` 作为 offsets，`cache_seqlens` 作为 `seq_len_decoder`

3. `test_radix_topk_accuracy.py`:
   - 更新 decode 模式测试用例，使用 `cu_seqlens_q` 替代 `offsets`

问题

级别	文件	概述
🟡 建议	indexer_topk.cuh:2061	注释语法错误（/ 应为 //）
❓ 疑问	indexer_topk.cuh:2068	decode 模式下 batch_id = bid / q_num_heads 与 cu_seqlens_q 的语义可能不匹配

总体评价

PR 的核心修改逻辑是正确的，通过修改 decode 模式下的输出指针计算方式（从基于 bid 改为基于 batch_id 的 aux_input[batch_id]），解决了 multi-batch 场景下的索引映射问题。删除冗余的 logits 重排逻辑是合理的优化。

但存在一个潜在的语义混淆问题：cu_seqlens_q 通常表示累积序列长度，而 kernel 中将其作为每个 batch 在输出中的偏移使用。虽然当前的实现可能在 DSA 的特定场景下工作正常，但建议在文档或注释中明确说明 cu_seqlens_q 在此处的具体语义。

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@1090f8b). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/models/deepseek_v3.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7252   +/-   ##
==========================================
  Coverage           ?   72.95%           
==========================================
  Files              ?      383           
  Lines              ?    53510           
  Branches           ?     8379           
==========================================
  Hits               ?    39039           
  Misses             ?    11746           
  Partials           ?     2725           

Flag	Coverage Δ
GPU	72.95% <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.