[BugFix] Fix save_output_specualate parameter bugs in suffix decoding

[Deleter-D]

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

proposer_share_inputs should be optional parameter since it's only needed for MTP prefill mode

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 | 2026-04-22 20:06:05

📋 Review 摘要

PR 概述：修复投机解码（Speculative Decoding）场景下 save_output 相关的两处 bug：draft_tokens 写入范围不足以及 proposer_share_inputs 在非 MTP 场景下的非法传参。
变更范围：spec_decode/base.py、model_executor/pre_and_post_process.py、worker/gpu_model_runner.py
影响面 Tag：[Speculative Decoding] [BugFix]

---

📝 PR 规范检查

PR 的 Motivation 和 Modifications 章节均未填写，无法了解 bug 的触发条件和修复思路，建议补充以下内容（可直接复制）：

描述建议（可直接复制）：

## Motivation
在投机解码（Speculative Decoding）MTP 模式下，存在以下两处 bug：
1. `prepare_dummy_speculative_drafts` 中 `draft_tokens` 写入范围为 `[:max_fake_drafts]`，
   但 `seq_lens_this_time` 被设置为 `max_fake_drafts + 1`，两者不一致导致 CUDAGraph 捕获时
   最后一列 token 未被正确初始化，可能引发推理异常。
2. `save_output_specualate` 在非 MTP prefill 场景下仍强制要求传入 `proposer_share_inputs`，
   导致非 MTP 路径调用时出错。

## Modifications
- `spec_decode/base.py`：将 `draft_tokens` 写入列索引从 `:max_fake_drafts` 修正为 `:max_fake_drafts + 1`，与 `seq_lens_this_time` 保持一致。
- `model_executor/pre_and_post_process.py`：将 `proposer_share_inputs` 改为可选参数（默认 `None`），
  并在 `is_mtp_prefill=True` 分支入口处增加 `assert` 保护。
- `worker/gpu_model_runner.py`：调用侧按 `spec_method == MTP` 条件传入 `proposer.model_inputs` 或 `None`。

---

问题

级别	文件	概述
🟡 建议	spec_decode/base.py:121	写入范围修复正确，建议补充注释说明 +1 的语义

---

总体评价

三处修改逻辑连贯、方向正确，proposer_share_inputs 的参数设计由强制传参改为按需传入，提升了接口健壮性；draft_tokens 写入范围修复消除了 CUDAGraph 场景下的潜在不一致风险。建议补全 PR 描述以便后续追溯。

[PaddlePaddle-bot]

🟡 建议 修复了 draft_tokens 写入范围与 seq_lens_this_time 不一致的 bug

修改前仅写入 [:max_fake_drafts] 列，但 seq_lens_this_time 设置为 max_fake_drafts + 1，导致最后一列 token 未初始化（仍为历史值），在 CUDAGraph 捕获时可能产生不一致行为。修复后写入范围与 seq_len 一致，逻辑正确。

建议补充注释说明 +1 的语义（1 个 target token 位 + max_fake_drafts 个 draft token 位）：

# draft_tokens layout: [target_token_placeholder, draft_token_0, ..., draft_token_{N-1}]
share_inputs["draft_tokens"][:batch_size, : max_fake_drafts + 1] = 5

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@b19754c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/pre_and_post_process.py	0.00%	1 Missing ⚠️
fastdeploy/spec_decode/base.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7566   +/-   ##
==========================================
  Coverage           ?   72.46%           
==========================================
  Files              ?      419           
  Lines              ?    57649           
  Branches           ?     9037           
==========================================
  Hits               ?    41773           
  Misses             ?    13058           
  Partials           ?     2818           

Flag	Coverage Δ
GPU	72.46% <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.