[BugFix] fix get_request from scheduler

[juncaipeng]

Motivation

修复v1调度下从 splitwise_scheduler 拉不到长请求的问题

v1调度设置scheduler_config.max_num_batched_tokens为8192

 if self.scheduler_config.max_num_batched_tokens is None:
            if int(envs.ENABLE_V1_KVCACHE_SCHEDULER):
                if paddle.is_compiled_with_xpu():
                    self.scheduler_config.max_num_batched_tokens = self.model_config.max_model_len
                else:
                    self.scheduler_config.max_num_batched_tokens = 8192  # if set to max_model_len, it's easy to be OOM
            else:
                if self.cache_config.enable_chunked_prefill:
                    self.scheduler_config.max_num_batched_tokens = 2048
                else:
                    self.scheduler_config.max_num_batched_tokens = self.model_config.max_model_len

基于scheduler_config.max_num_batched_tokens来从 splitwise_scheduler 拉取请求

  if self.cfg.scheduler_config.splitwise_role != "mixed":
      max_num_batched_tokens = self.cfg.scheduler_config.max_num_batched_tokens
  else:
      max_num_batched_tokens = self.cfg.model_config.max_model_len

tasks = self.scheduler.get_requests(
                    available_blocks=available_blocks,
                    block_size=self.cfg.cache_config.block_size,
                    reserved_output_blocks=self.cfg.cache_config.enc_dec_block_num,
                    max_num_batched_tokens=max_num_batched_tokens,
                    batch=num_prefill_batch,
                )

Modifications

修改 splitwise_scheduler

Usage or Command

无

Accuracy Tests

无

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.

[Copilot]

Pull request overview

This PR fixes a critical bug in the v1 scheduler where long requests (those exceeding max_num_batched_tokens) could not be processed from the InferScheduler, causing an infinite loop.

Key Changes

• Added len(reqs) > 0 condition to line 862 in splitwise_scheduler.py to allow the first request to be processed even when it exceeds max_num_batched_tokens
• This prevents the scheduler from repeatedly pushing back and re-attempting to process the same long request

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@3697110). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/scheduler/splitwise_scheduler.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5369   +/-   ##
==========================================
  Coverage           ?   59.79%           
==========================================
  Files              ?      325           
  Lines              ?    40286           
  Branches           ?     6100           
==========================================
  Hits               ?    24090           
  Misses             ?    14294           
  Partials           ?     1902           

Flag	Coverage Δ
GPU	59.79% <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.

[paddle-bot]

Thanks for your contribution!