[XPU] add verify draft tokens

[cmcamdy]

add verify draft tokens
#6685

Motivation

💡 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

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!

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@ea998dd). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6947   +/-   ##
==========================================
  Coverage           ?   73.51%           
==========================================
  Files              ?      383           
  Lines              ?    53614           
  Branches           ?     8412           
==========================================
  Hits               ?    39417           
  Misses             ?    11516           
  Partials           ?     2681           

Flag	Coverage Δ
GPU	73.51% <ø> (?)

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

[Copilot]

Pull request overview

该 PR 在 XPU（Kunlun3）侧新增 verify_draft_tokens 的插件/算子实现，并通过 pybind 暴露到 Python，同时补充了对应的单测，用于投机解码（Speculative Decoding）中对 draft tokens 的校验与 Phase2 采样输出。

Changes:

• 新增 XPU3 kernel：verify_draft_tokens.xpu，实现 Phase1 verify + Phase2 输出逻辑
• 新增插件 wrapper + Paddle 扩展算子 + pybind 导出：对外提供 verify_draft_tokens 调用入口
• 新增较完整的 Python 单测：覆盖不同 verify_strategy、verify_window、EOS/max_dec_len、skip 条件等

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
custom_ops/xpu_ops/test/test_verify_draft_tokens.py	新增 verify_draft_tokens 的 Python 单测与参考实现
custom_ops/xpu_ops/src/plugin/src/wrapper/mtp_wrapper/verify_draft_tokens.cpp	新增 verify_draft_tokens 插件 wrapper（CPU/XPU3 wrapper + 参数校验）
custom_ops/xpu_ops/src/plugin/src/kernel/kunlun3cpp/mtp_kernel/verify_draft_tokens.xpu	新增 Kunlun3 XPU kernel 实现
custom_ops/xpu_ops/src/plugin/include/xpu/plugin.h	新增 verify_draft_tokens 插件函数声明
custom_ops/xpu_ops/src/ops/pybind/pybind.cc	新增 pybind 导出 verify_draft_tokens
custom_ops/xpu_ops/src/ops/mtp/verify_draft_token.cc	新增 Paddle 扩展算子 host 侧实现与参数校验
custom_ops/xpu_ops/src/ops/mtp/speculate_verify.cc	调整 speculate_verify 的随机数生成逻辑（持久 seed/offset）
custom_ops/xpu_ops/src/plugin/src/wrapper/mtp_wrapper/speculate_verify.cpp	清理重复 typedef 声明

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-14

📋 Review 摘要

PR 概述：为 XPU 硬件添加 verify_draft_tokens 算子，支持 TOPP/GREEDY/TARGET_MATCH 三种验证策略，用于 speculative decoding 的 draft token 验证。

变更范围：custom_ops/xpu_ops/ (新增 7 个文件，修改 1 个文件)

• 新增 verify_draft_token.cc (host 端实现)
• 新增 verify_draft_tokens.xpu (XPU kernel)
• 新增 verify_draft_tokens.cpp (wrapper)
• 新增 test_verify_draft_tokens.py (单元测试)
• 修改 speculate_verify.cc (添加 persistent seed/offset)

影响面 Tag：[XPU] [OP]

📝 PR 规范检查

PR 描述未遵循官方模板，缺少以下必填章节：

标题建议（当前已有有效 tag）：

• 当前标题：[XPU] add verify draft tokens ✓

描述建议（请补充以下内容）：

## Motivation

为 XPU 硬件添加 draft token 验证算子，支持 speculative decoding 场景。参考 PR #6685 实现 XPU 版本。

## Modifications

1. 新增 `verify_draft_token.cc`：算子 host 端实现，支持三种验证策略
   - TOPP (0)：验证 draft token 是否在 top-p 候选集中
   - GREEDY (1)：验证 draft token 是否等于 argmax token
   - TARGET_MATCH (2)：验证 draft token 是否等于目标模型的采样 token

2. 新增 `verify_draft_tokens.xpu`：XPU3 kernel 实现，包含 VerifyContext 结构体和验证逻辑

3. 新增 `verify_draft_tokens.cpp`：CPU/XPU3 wrapper 实现

4. 新增 `test_verify_draft_tokens.py`：单元测试（包含 Python 参考实现）

5. 修改 `speculate_verify.cc`：使用 persistent seed/offset 替代固定 random seed

## Usage or Command

```python
from fastdeploy.model_executor.ops.xpu import verify_draft_tokens

verify_draft_tokens(
    step_output_ids,      # 输出 token ids
    step_output_len,       # 输出长度
    step_input_ids,       # 输入 draft tokens
    target_tokens,        # target 模型输出 (GREEDY/TARGET_MATCH)
    candidate_ids,        # 候选 token ids (TOPP/GREEDY)
    candidate_scores,     # 候选分数 (TOPP)
    candidate_lens,      # 候选长度 (TOPP/GREEDY)
    topp,               # top-p 阈值
    stop_flags,         # 停止标志
    seq_lens_encoder,    # encoder 序列长度
    seq_lens_this_time, # 当前序列长度
    end_tokens,         # EOS tokens
    is_block_step,      # 是否 block step
    cu_seqlens_q_output,# 累积序列长度
    reasoning_status,    # 推理状态
    max_dec_len,       # 最大解码长度
    step_idx,          # 当前步索引
    max_seq_len=4096,
    verify_window=4,
    verify_strategy=0,  # 0=TOPP, 1=GREEDY, 2=TARGET_MATCH
    reject_all=False,
    accept_all=False
)

Accuracy Tests

已添加单元测试 test_verify_draft_tokens.py，包含 Python 参考实现与 XPU 实现的对比测试。

Checklist

• Add at least a tag in the PR title.
• Format your code, run pre-commit before commit. (CI PASSED)
• Add unit tests. Added test_verify_draft_tokens.py with reference implementation.
• Provide accuracy results. (待补充)
• If the current PR is submitting to the release branch... (N/A)

### 问题

| 级别 | 文件 | 概述 |
|------|------|------|
| 🟡 建议 | - | PR 描述不完整，缺少 Motivation/Modifications/Usage/Accuracy Tests 章节 |

### 总体评价

代码实现逻辑正确，包含完整的 Python 参考实现和单元测试。VerifyContext 结构体在 XPU kernel 和 wrapper 中实现一致，emit_token/emit_final_token 的 EOS/max_dec_len 检测逻辑正确。使用 persistent seed/offset 确保随机数质量与 GPU 版本对齐。主要问题是 PR 描述未遵循模板，请补充 Motivation/Modifications/Usage/Accuracy Tests 内容。