[Optim] Robust sync status when preempted happens

[rainyfly]

Motivation

背景

当抢占发生时，服务层需要在发生抢占的槽位和引擎做一次同步，确保该槽位引擎生成的 token 都已经接收完毕，才可以被调度回 waiting 队列进行重调度。
如果服务层发动了对请求的抢占，却没有和引擎做同步，就可能导致服务层将被抢占的请求调度回 prefill时候看到的 need prefill token 数是 N，而之后发给引擎时候又由于异步接收到了额外一个 token，最终引擎层看到的need prefill token 确是 N+1，这种不同步会导致该槽位的请求 hang 住。
为了解决这一同步问题，目前服务层的 token processor 通过 token_id -1来保证同步，逻辑是引擎在吐完有效 token 后，处理 preempted task 时会返回-1，代表之前该槽位生成的有效 token 都已经接收完毕，之后再把被抢占请求放回 waiting 队列以便被重调度。

但是实际上，生成-1 会代表两个含义：

1. 该槽位没有请求，即本身无效槽位的地方会生成-1
2. 该槽位上已分配的 block资源已经填充完毕，或者 chunked prefill 时候，也会生成-1

用-1来做抢占时候的同步，在某些场景下可能会无法判断是情况 1 还是情况 2 所生成的-1 ，从而带来同步问题。
为了增强鲁棒性并且避免歧义，采用单独的 token_id 用做同步引擎对抢占的处理。

Modifications

None

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!

[Copilot]

Pull request overview

This PR fixes a synchronization issue that occurs during request preemption by introducing a dedicated PREEMPTED_TOKEN_ID (-9) to distinguish preemption synchronization signals from other uses of token_id -1. Previously, -1 was ambiguously used for both invalid slots and completed block allocations, which could cause synchronization problems when the token processor processes tokens slower than the engine generates them.

Key changes:

• Introduces PREEMPTED_TOKEN_ID = -9 constant to explicitly signal preemption completion
• Adds preempted_idx list tracking in all model runners (GPU, XPU, Metax) to record which slots have been preempted
• Updates token processor logic to specifically check for PREEMPTED_TOKEN_ID before rescheduling preempted requests
• Adds guard in cache_output_tokens to only cache during decoding phase, preventing issues when requests are preempted during prefill

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
fastdeploy/config.py	Defines new PREEMPTED_TOKEN_ID = -9 constant for unambiguous preemption signaling
fastdeploy/worker/xpu_model_runner.py	Initializes and tracks preempted indices in share_inputs for XPU model runner
fastdeploy/worker/metax_model_runner.py	Initializes and tracks preempted indices in share_inputs for Metax model runner
fastdeploy/worker/gpu_model_runner.py	Initializes and tracks preempted indices in share_inputs for GPU model runner
fastdeploy/output/token_processor.py	Updates reschedule logic to check for specific PREEMPTED_TOKEN_ID; removes old batch-based reschedule method; adjusts token_id comparison from <= 0 to < 0
fastdeploy/model_executor/pre_and_post_process.py	Sets PREEMPTED_TOKEN_ID for preempted slots after token generation and input updates
fastdeploy/engine/sched/resource_manager_v1.py	Adds condition to only cache output tokens during decoding phase, preventing issues during preemption in prefill

[codecov-commenter]

Codecov Report

❌ Patch coverage is 60.00000% with 8 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@5afeef6). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/output/token_processor.py	63.63%	0 Missing and 4 partials ⚠️
fastdeploy/engine/sched/resource_manager_v1.py	0.00%	2 Missing ⚠️
...tdeploy/model_executor/xpu_pre_and_post_process.py	0.00%	1 Missing ⚠️
fastdeploy/worker/gpu_model_runner.py	75.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5796   +/-   ##
==========================================
  Coverage           ?   67.30%           
==========================================
  Files              ?      348           
  Lines              ?    44769           
  Branches           ?     6891           
==========================================
  Hits               ?    30134           
  Misses             ?    12409           
  Partials           ?     2226           

Flag	Coverage Δ
GPU	67.30% <60.00%> (?)

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.