[Cherry-Pick][BugFix][PD Disaggregation] remove redundant block allocation of prefill tasks in decode instance (#8022)

[liyonghua0910]

Motivation

In the disaggregated prefill-decode serving mode (splitwise_role != "mixed"), the decode instance incorrectly applied the mixed-role block reservation logic when determining whether to admit a new prefill request. This caused the scheduler to over-provision KV-cache blocks for running decode requests, unnecessarily blocking prefill task admission and reducing throughput. Additionally, get_new_block_nums could return a negative block count when a request already had sufficient blocks allocated, potentially leading to downstream scheduling errors.

Modifications

• fastdeploy/engine/sched/resource_manager_v1.py
  1. _get_can_schedule_prefill_threshold_block: Return num_chunk_new_block directly when enabling prefill-decode disaggregation, skipping the reserve-block estimation for running requests.
  2. get_new_block_nums: Clamp the result to non-negative, preventing negative values when num_computed_tokens + num_new_tokens fits within already-allocated blocks.

Usage or Command

No new configuration or command required. The fix applies automatically in disaggregated prefill-decode deployments.

Accuracy Tests

This PR only modifies scheduling logic (block allocation thresholds) and does not affect model forward computation or kernel outputs, so accuracy is not impacted.

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

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (release/2.6@a869a06). Learn more about missing BASE report.

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.6    #8021   +/-   ##
==============================================
  Coverage               ?   71.55%           
==============================================
  Files                  ?      386           
  Lines                  ?    55740           
  Branches               ?     8753           
==============================================
  Hits                   ?    39886           
  Misses                 ?    13034           
  Partials               ?     2820           

Flag	Coverage Δ
GPU	71.55% <100.00%> (?)

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 | pr_review | 2026-06-08 21:37:20

📋 Review 摘要

PR 概述：调整 ResourceManagerV1 的 prefill block 计算与调度阈值，避免 PD 分离 decode 实例为 prefill 任务重复预留/申请 block。
变更范围：fastdeploy/engine/sched/resource_manager_v1.py
影响面 Tag：[Scheduler] [PD Disaggregation]

问题

级别	文件	概述
-	-	未发现阻塞性问题。PR 规范问题在下面章节报，不要在这里重复。

📝 PR 规范检查

不符合规范：目标分支是 release/2.6，但标题未使用 release PR 要求的 Cherry-Pick 格式；同时 PR 描述中的 Motivation、Modifications、Usage or Command、Accuracy Tests 仍为空模板，Checklist 也未按实际情况勾选。当前 PR 信息和提交信息都未提供原始 develop PR ID，因此标题中的原 PR 号需要作者补充。

标题建议（补充原始 develop PR ID 后可直接复制）：

• [Cherry-Pick][BugFix] remove redundant block allocation of prefill tasks in decode instance(#原始develop PR号)

PR 描述建议（点击展开，可直接复制）

## Motivation
Fix redundant KV-cache block reservation/allocation for prefill tasks on decode instances in PD disaggregation. Decode-only instances should not apply the mixed-role prefill reserve threshold when checking whether a prefill-related task can be admitted.

## Modifications
- `fastdeploy/engine/sched/resource_manager_v1.py`: clamp `get_new_block_nums()` after speculative block adjustment so over-allocated requests do not request another redundant block.
- `fastdeploy/engine/sched/resource_manager_v1.py`: limit `_get_can_schedule_prefill_threshold_block()` reserve logic to `splitwise_role == "mixed"`; prefill/decode roles use the actual new block count.
- `fastdeploy/engine/sched/resource_manager_v1.py`: unify running prefill scheduling threshold calculation through `_get_can_schedule_prefill_threshold_block()`.

## 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.

总体评价

调度逻辑修改范围较小，代码层面看起来与避免 decode 角色冗余 prefill block reserve 的目标一致。合入前请补齐 release Cherry-Pick 元信息和 PR 描述，便于维护者确认来源与验证范围。