[XPU] Split the block_attn operator into smaller operators

[RuohengMa]

Motivation

splice block_attn fused op

Modifications

splice block_attn

Usage or Command

export encoder_splice=1 to enable spliced split_rope/neox_cache_kv_encoder
export decoder_splice=1 to enable spliced split_rope/neox_cache_kv_decoder

Accuracy Tests

None

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

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@26d6a20). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6798   +/-   ##
==========================================
  Coverage           ?   74.47%           
==========================================
  Files              ?      383           
  Lines              ?    53619           
  Branches           ?     8412           
==========================================
  Hits               ?    39931           
  Misses             ?    10968           
  Partials           ?     2720           

Flag	Coverage Δ
GPU	74.46% <100.00%> (?)

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.

[hong19860320]

LGTM

[iosmers]

LGTM

[PaddlePaddle-bot]

📋 Review 摘要

PR 概述：为 XPU 平台引入 splice 方式的 block_attn 算子，支持通过环境变量 encoder_splice 和 decoder_splice 开启 splice 模式。

变更范围：custom_ops/xpu_ops/、fastdeploy/model_executor/layers/backends/xpu/、fastdeploy/worker/、fastdeploy/spec_decode/

影响面 Tag：[XPU] [OP] [Speculative Decoding]

📝 PR 规范检查

PR 描述基本符合要求，但 Checklist 存在以下问题：

项目	状态	说明
Add at least a tag in PR title	✅	已包含 [XPU] tag
Format your code, run pre-commit	❓	未勾选，建议确认是否已执行
Add unit tests	❌	未勾选，但实际新增了测试文件，建议补充单元测试覆盖说明
Provide accuracy results	❌	未提供准确性验证结果
If submitting to release branch	✅	当前提交到 develop 分支

标题建议：当前标题 [XPU] splice block_attn 符合规范，建议可考虑更明确地描述优化类型：

• [Optimization][XPU] Add spliced block_attn for better performance

问题

级别	文件	概述
🟡 建议	custom_ops/xpu_ops/download_dependencies.sh:18	依赖版本改为 latest 可能导致构建不稳定性
🟡 建议	docs/	缺少新功能的使用文档（encoder_splice/decoder_splice）
🟡 建议	custom_ops/xpu_ops/test/test_block_attn.py	测试存在多个 TODO 和被注释的用例
🟡 建议	PR 描述	缺少性能对比数据和准确性验证结果

总体评价

代码实现整体结构合理，splice 方式的实现思路清晰。但存在以下待改进之处：

1. 依赖管理：将固定版本改为 latest 可能导致不同时间构建产生不同结果，影响可复现性
2. 文档缺失：新增的 splice 功能缺少使用文档，用户无法了解如何正确使用
3. 测试覆盖：测试代码中存在多处 TODO 和注释掉的测试用例（如 mixed mode、mtp branch），表明功能尚未完全验证
4. 缺少数据：未提供性能对比数据，无法验证 splice 模式是否真正带来性能提升

建议补充文档和测试覆盖后再合并，或作为实验性功能提供明确的标识。

[PaddlePaddle-bot]

🟡 建议 将依赖版本从固定版本号改为 latest 可能导致构建不稳定性，建议使用具体的版本号或通过配置管理。

原因：使用 latest 标签会导致不同时间拉取到不同版本的依赖，影响构建的可复现性和稳定性。

建议修复方式：

# 使用明确的版本号
version_xvllm="20260415"  # 或其他明确版本

[yongqiangma]

LGTM

[gongshaotian]

LGTM for CUDAGraph