[Optimize] Optimize ttft for ep

[rainyfly]

Motivation

优化 zmq -> scheduler -> worker 的调度逻辑。
之前scheduler 和 worker 是完全异步分开进行的，schedule()会在引擎forward开始前发生并立即结束，容易造成每次调度的请求量不满，从而让请求易发生排队。
实际上，在整个引擎 forward 期间，都是可以预处理请求并且接收请求放在scheduler的队列里累积，在引擎 forward 结束后再进行schedule()，此时才能达到最佳组batch的效果。

Modifications

1. 引擎 forward结束再 schedule()
2. forward期间持续接收请求并做预处理，放到 scheduler的队列中等待 schedule()。

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!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 53.84615% with 24 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@9b0a82c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/engine/common_engine.py	14.28%	12 Missing ⚠️
fastdeploy/worker/worker_process.py	64.28%	6 Missing and 4 partials ⚠️
fastdeploy/scheduler/dp_scheduler.py	85.71%	0 Missing and 1 partial ⚠️
fastdeploy/splitwise/internal_adapter_utils.py	66.66%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6098   +/-   ##
==========================================
  Coverage           ?   67.66%           
==========================================
  Files              ?      389           
  Lines              ?    51915           
  Branches           ?     8080           
==========================================
  Hits               ?    35130           
  Misses             ?    14234           
  Partials           ?     2551           

Flag	Coverage Δ
GPU	67.66% <53.84%> (?)

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.

[Copilot]

Pull request overview

此 PR 优化了专家并行（EP）场景下的调度逻辑，通过将 schedule() 调用时机从引擎 forward 开始前移到 forward 结束后，以实现更好的请求批处理效果。主要目标是优化 Time To First Token (TTFT)。

Changes:

• 引入 engine_forward_signal IPC 信号来协调调度器和 worker 之间的同步
• 简化 DPLocalScheduler.get_requests() 方法，移除资源检查和超时逻辑
• 调整 worker 事件循环，将 EP 负载均衡（eplb）操作移至引擎 forward 之后
• 修改调度器循环以等待引擎 forward 完成后再进行调度
• 为 EP 场景添加空任务处理机制

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
fastdeploy/scheduler/dp_scheduler.py	大幅简化 get_requests 方法，移除资源检查、批处理和超时逻辑；删除 envs 导入
fastdeploy/engine/common_engine.py	添加 engine_forward_signal 初始化；修改调度循环以等待 forward 完成；为 EP 添加空任务处理
fastdeploy/worker/worker_process.py	添加 engine_forward_signal 支持；在检测到任务和完成 forward 时设置/重置信号；添加 EP prefill barrier；移动 eplb 调用时机
tests/scheduler/test_dp_scheduler.py	删除两个测试资源不足场景的测试用例
tests/ci_use/metrics/test_metrics.py	注释掉验证 clear_load_weight 后指标为 0 的断言

Comments suppressed due to low confidence (1)

tests/scheduler/test_dp_scheduler.py:478

• 现有的测试用例 test_get_requests_timeout 和 test_get_requests_no_requests_available 现在已经过时，因为它们仍然 mock envs.FD_EP_BATCHED_TOKEN_TIMEOUT，但新的实现不再使用超时逻辑。

这些测试现在可能会产生误导性结果或给出假阳性。建议：

1. 更新这些测试以反映新的实现行为
2. 或者删除这些测试，如果它们测试的场景不再适用
3. 验证 test_get_requests_successful_batching 仍然正确测试新的行为

保留使用已删除功能的测试会让未来的维护者困惑。

    @patch("time.time")
    @patch.object(dp_scheduler_module, "envs")
    def test_get_requests_no_requests_available(self, mock_envs, mock_time):
        """Test getting requests when no requests are available."""
        # Mock envs to return our mock environment
        mock_envs.FD_EP_BATCHED_TOKEN_TIMEOUT = 0.1

        # Mock time to return consistent values - provide enough values for multiple calls
        time_values = [100.0, 100.1, 100.2, 100.3, 100.4, 100.5]  # Multiple values for the loop
        mock_time.side_effect = time_values

        # Mock the condition variable to avoid threading issues
        with patch.object(self.scheduler, "requests_not_empty"):
            requests = self.scheduler.get_requests(
                available_blocks=20, block_size=16, reserved_output_blocks=10, max_num_batched_tokens=1024, batch=1
            )

        # Should return empty list after timeout
        self.assertEqual(requests, [])

    def test_get_requests_successful_batching(self):
        """Test successful request batching."""
        # Add a mock request
        mock_request = MockRequest("test_req", prompt_tokens_ids_len=10)
        self.scheduler.requests["test_req"] = mock_request
        self.scheduler.ids = ["test_req"]

        # Mock calc_required_blocks to return small value
        self.scheduler.calc_required_blocks = Mock(return_value=1)

        requests = self.scheduler.get_requests(
            available_blocks=20, block_size=16, reserved_output_blocks=10, max_num_batched_tokens=1024, batch=1
        )

        # Should get the request
        self.assertEqual(len(requests), 1)
        self.assertEqual(requests[0].request_id, "test_req")

    @patch("time.time")
    @patch.object(dp_scheduler_module, "envs")
    def test_get_requests_timeout(self, mock_envs, mock_time):
        """Test request batching with timeout."""
        # Mock envs to return our mock environment
        mock_envs.FD_EP_BATCHED_TOKEN_TIMEOUT = 0.1

        # Mock time to return consistent values - provide enough values for multiple calls
        time_values = [100.0, 100.1, 100.2, 100.3, 100.4, 100.5]  # Multiple values for the loop
        mock_time.side_effect = time_values

        # Add a mock request
        mock_request = MockRequest("test_req", prompt_tokens_ids_len=10)
        self.scheduler.requests["test_req"] = mock_request
        self.scheduler.ids = ["test_req"]

        # Mock calc_required_blocks to return large value to exceed available blocks
        self.scheduler.calc_required_blocks = Mock(return_value=50)

        # Mock the condition variable to avoid threading issues
        with patch.object(self.scheduler, "requests_not_empty"):
            requests = self.scheduler.get_requests(
                available_blocks=20, block_size=16, reserved_output_blocks=10, max_num_batched_tokens=1024, batch=1
            )

        # Should return empty due to timeout
        self.assertEqual(requests, [])