[XPU] get_infer_param use inplace copy, remove block_tables abundant d2h copy

[RuohengMa]

Motivation

get_infer_param use inplace copy, remove block_tables abundant d2h copy

Modifications

get_infer_param use inplace copy, remove block_tables abundant d2h copy

Usage or Command

None

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

❌ Patch coverage is 4.34783% with 22 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@91b8bf2). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/forward_meta.py	4.34%	22 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7431   +/-   ##
==========================================
  Coverage           ?   73.84%           
==========================================
  Files              ?      398           
  Lines              ?    55000           
  Branches           ?     8613           
==========================================
  Hits               ?    40614           
  Misses             ?    11671           
  Partials           ?     2715           

Flag	Coverage Δ
GPU	73.84% <4.34%> (?)

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]

🤖 AI Code Review | 2026-04-17 17:45 CST

📋 Review 摘要

PR 概述：将 get_infer_param 算子改为 inplace copy 模式，消除 block_tables 冗余的 D2H 拷贝；同时修复 block_attn_spliced 在 rope_3d 场景下的分支问题。
变更范围：custom_ops/xpu_ops/（C++ 算子实现 + 测试）、fastdeploy/model_executor/（Python 调用侧 + ForwardMeta）
影响面 Tag：XPU OP Optimization

📝 PR 规范检查

PR 标题已包含 [XPU] tag，符合规范。但 Modifications 部分与 Motivation 完全相同，建议补充具体的修改点说明。此外 Checklist 项目均未勾选。

描述建议（可直接复制）：

Modifications

1. get_infer_param 算子改为 inplace 模式：将 22 个输出 tensor 改为输入+inplace 输出，避免每次调用重新分配内存
2. lod_to_slot_mapping 中去除 block_tables 的 D2H 拷贝，改用 index_select 在 XPU 上直接完成 slot mapping 计算
3. block_attn_spliced 在 rope_3d=true 时跳过 splice 路径，避免触发 PD_THROW
4. Python 侧 xpu_pre_and_post_process 支持 cudagraph 模式下的 inplace copy

问题

级别	文件	概述
🟡 建议	block_attn_spliced.cc:244	外层加了 !rope_3d 后，内层 if (rope_3d) 成为死代码
❓ 疑问	get_infer_param.cc:221	循环上界从 bsz 改为 bsz + 1，需确认变更意图

总体评价

核心优化思路清晰：通过 inplace copy 避免重复分配 tensor，通过 index_select 替代 D2H+CPU 计算消除 block_tables 的冗余拷贝，对性能有明显帮助。代码逻辑正确，测试文件均已同步更新。两个小问题建议关注：block_attn_spliced.cc 中的死代码清理，以及 decoder_seq_lod_vec 循环上界变更的意图说明。

[PaddlePaddle-bot]

🟡 建议 外层条件已加上 !rope_3d 后，内层 if (rope_3d) 永远不会为 true，形成死代码。encoder（此处）和 decoder（line 553）两处均存在相同问题。

建议清理内层的 rope_3d 检查和 PD_THROW，或在 TODO 注释中说明保留原因：

// TODO: spliced split kvcache should support rope3d
if (FLAGS_encoder_splice && !rope_3d) {
    // 以下 if (rope_3d) 已不可达，建议移除

[PaddlePaddle-bot]

❓ 疑问 循环上界从 bsz 改为 bsz + 1。decoder_seq_lod_vec 大小为 bsz + 1，访问不会越界，逻辑上也需要处理最后一个 lod 值。

想确认：这是修复了之前遗漏最后一个元素导致 monotonic 填充不完整的 bug，还是 cudagraph 模式下有新的需求？建议在注释中补充变更原因，方便后续维护。