[Iluvatar] Refactor transpose and reverse_transpose

[wuyujiji]

Motivation

修复prefill和decode mxied attention的bug：如果prefill 的本次长度等于 1，但这里把 seq_len == 1 一律当 decode 处理了，导致在reorder prefill和decode hidden state时shape没对上

Modifications

N/A

Usage or Command

N/A

Accuracy Tests

N/A

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.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 14 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@165f827). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/iluvatar_model_runner.py	0.00%	8 Missing ⚠️
...rs/backends/iluvatar/attention/mha_attn_backend.py	0.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8065   +/-   ##
==========================================
  Coverage           ?   67.51%           
==========================================
  Files              ?      475           
  Lines              ?    66879           
  Branches           ?    10312           
==========================================
  Hits               ?    45152           
  Misses             ?    18860           
  Partials           ?     2867           

Flag	Coverage Δ
GPU	77.53% <ø> (?)
XPU	6.95% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 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]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-06-22 14:16:10 UTC+08:00

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: c008c42 | Merge base: 165f827 (branch: develop)

---

1 Required任务 : 9/10 通过

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
42(0)	42	39	3	0	0	0

任务	错误类型	置信度	日志
Approval	需要 Approval	高	Job

2 失败详情

🔴 Approval — 需要 Approval（置信度: 高）

该 Job 需要人工 Approval，完成审批后 CI 才会继续执行。

修复建议：请通过人工审批。

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-23 11:17:45

📋 Review 摘要

PR 概述：重构 Iluvatar mixed attention hidden states transpose/reverse_transpose，新增设备端 reorder custom op，并调整示例参数。
变更范围：custom_ops/iluvatar_ops/、Iluvatar attention backend/model runner、Iluvatar 安装文档
影响面 Tag：OP Iluvatar Docs

问题

级别	文件	概述
🟡 建议	custom_ops/iluvatar_ops/mixed_reorder_hidden_states.cu:97	shape/offset 索引统一使用 32 位 int，缺少范围校验，超大 token/hidden 配置下可能溢出

历史 Findings 修复情况

Finding	问题	状态
F1	is_decode 时强制 copy_seq_len = 1 会丢掉 speculative/MTP decode 的额外 token。	⚠️ 仍存在

📝 PR 规范检查

标题 Tag [Iluvatar] 合规，但 PR 描述中 Modifications 仍为 N/A，且本次新增 custom op 未看到对应单测/精度说明。建议使用下面可直接复制的标题和描述。

标题建议（可直接复制）：

• [Iluvatar] Fix mixed attention hidden state reorder

PR 描述建议（点击展开，可直接复制）

## Motivation
修复 Iluvatar mixed attention 中 prefill 与 decode 混合时 hidden states 重排错误：prefill 本次长度为 1 时仍应按 seq_lens_encoder > 0 识别为 prefill，避免 reorder/reverse_reorder 后 shape 与顺序不匹配。

## Modifications
- 新增 `custom_ops/iluvatar_ops/mixed_reorder_hidden_states.cu`，按 `seq_lens_encoder`、`seq_lens_decoder`、`seq_lens_this_time` 在设备端重排 mixed attention hidden states。
- 在 Iluvatar MHA backend 中接入 `mixed_reorder_hidden_states`，并将 decode batch 判定改为 `seq_lens_encoder == 0 && seq_lens_decoder > 0`。
- 将新增 Iluvatar custom op 加入 `custom_ops/setup_ops.py` 编译列表。
- 更新 Iluvatar GPU 安装文档中 PaddleOCR-VL/PaddleOCR-VL-1.6 的示例服务参数。

## 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.
- [x] 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.
- [x] 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.

总体评价

本轮重点审查了新增 Iluvatar reorder custom op、mixed attention 调用链和 runner metadata 初始化。当前仍有一个历史 Bug 未修复，另有新增 custom op 索引类型需要补齐边界保护；建议在修复后补充覆盖 prefill length = 1、mixed prefill/decode 以及 decode 多 token 的回归测试。

[PaddlePaddle-bot]

🟡 建议 新增重排 kernel 把 hidden_shape[1] 和 seq_lens_this_time.shape()[0] 收窄成 int，后续 copy_seq_len * hidden_dim、original_offset、reordered_offset 也全部用 32 位索引计算。

这些值来自 tensor shape 和调度 token 数，当前没有任何 <= INT_MAX / 乘法溢出校验；一旦 max_num_batched_tokens * hidden_dim 或 offset 超过 32 位范围，索引会回绕并读写错误位置。FastDeploy checklist 对 custom op 的 shape/offset 承接要求使用 int64_t。

建议修复方式：
将 host 侧 hidden_dim/max_num_seqs 和 kernel 内 original_start、reordered_start、idx、offset 相关变量改为 int64_t 计算；如果 Iluvatar kernel 必须使用 32 位索引，则在 launch 前显式校验 hidden_shape[0] * hidden_shape[1] <= INT_MAX、hidden_shape[1] <= INT_MAX 并给出清晰报错。