[PD Disaggregation] remove splitwise deployment on single node and refine the code

[juncaipeng]

Motivation

remove splitwise deployment on single node and refine the code

Modifications

• remove the code, example and doc of splitwise deployment on single node
• splitwise deployment support v1 scheduler (must set --num-gpu-blocks-override in prefill)
• refine the function in fastdeploy/engine/common_engine.py such as _decode_process_splitwise_requests and _insert_prefilled_requests
• add test for splitwise deployment with splitwise_scheduler

Usage or Command

Refer to examples.

Accuracy Tests

Use ci to test accuracy.

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 refactors the splitwise (Prefill-Decode disaggregated) deployment architecture by removing the deprecated single-machine deployment mode (v2) and consolidating to two supported methods: v0 using splitwise_scheduler/dp_scheduler, and v1 using local_scheduler with router. The changes simplify the codebase by removing redundant code paths and improving the separation of concerns between prefill and decode instances.

• Removed deprecated innode_prefill_ports parameter and associated single-machine PD logic
• Refactored decode instance's request processing into a cleaner _decode_process_splitwise_requests function
• Updated test utilities to share common functions and removed duplicated port/service management code
• Consolidated splitwise version naming from v0/v1/v2 to just v0/v1

Reviewed Changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
tests/e2e/utils/serving_utils.py	Added shared utility functions check_service_health and get_registered_number for e2e tests
tests/e2e/test_ernie_03b_pd_splitwise_scheduler.py	Updated to use shared utilities, added redis support, removed duplicated helper functions
tests/e2e/test_ernie_03b_pd_router_v0.py	Refactored to import utilities from serving_utils instead of duplicating code
fastdeploy/worker/worker_process.py	Removed automatic ENABLE_V1_KVCACHE_SCHEDULER=0 setting for non-RDMA splitwise
fastdeploy/splitwise/splitwise_connector.py	Removed deprecated innode dispatch methods (has_splitwise_tasks, dispatch_innode_splitwise_tasks)
fastdeploy/output/token_processor.py	Added timeout warning for cache sending, removed v0-specific result filtering
fastdeploy/inter_communicator/engine_worker_queue.py	Removed available_prefill_instances queue that was used for single-machine coordination
fastdeploy/engine/request.py	Added timestamp fields inference_start_time and llm_engine_recv_req_timestamp to Request class
fastdeploy/engine/engine.py	Removed initialization of available_prefill_instances queue
fastdeploy/engine/common_engine.py	Major refactoring: extracted _insert_prefilled_requests, renamed _process_splitwise_task to _decode_process_splitwise_requests with cleaner logic
fastdeploy/engine/async_llm.py	Removed available_prefill_instances queue initialization
fastdeploy/engine/args_utils.py	Removed innode_prefill_ports argument, added validation for splitwise configuration
fastdeploy/demo/offline_disaggregated_demo.py	Deleted deprecated single-machine offline demo
fastdeploy/config.py	Simplified splitwise version detection logic (v0/v1 only), removed innode_prefill_ports config
fastdeploy/cache_manager/transfer_factory/ipc_cache_transfer.py	Removed unused finish_event variable
fastdeploy/cache_manager/cache_messager.py	Fixed connection status logging logic
examples/splitwise/start_v2_tp1.sh	Removed deprecated v2 single-machine example script
examples/splitwise/start_v1_tp*.sh	Updated scripts to use router-based v1 deployment method
examples/splitwise/start_v0_tp*.sh	Updated scripts to use splitwise_scheduler for v0 deployment
examples/splitwise/start_mixed.sh	Added test request example for mixed server deployment
docs/zh/features/disaggregated.md	Removed single-machine deployment documentation, updated multi-machine examples
docs/features/disaggregated.md	Removed single-machine deployment documentation, updated multi-machine examples

Comments suppressed due to low confidence (4)

tests/e2e/test_ernie_03b_pd_splitwise_scheduler.py:337

• Unused variable assignment. The variable p_url is assigned on line 336 but the function sends the request to p_url instead of d_url on line 337. If the intention was to test both URLs, clarify with a comment. If only prefill URL is needed, remove the d_url assignment.
  tests/e2e/test_ernie_03b_pd_splitwise_scheduler.py:366
• Inconsistent variable usage pattern. Lines 365-366 and 416-417 follow the pattern of unpacking both URLs but only using one. Similarly, lines 389-390 do the same. For better clarity, consider using tuple unpacking with underscore for unused values: p_url, _ = api_url when only using one URL.
  tests/e2e/test_ernie_03b_pd_splitwise_scheduler.py:164
• Typo in variable assignment. There's a typo on line 164: env_prefill is incorrectly used instead of env_decode. This should be env_decode["ENABLE_V1_KVCACHE_SCHEDULER"] = "0" to properly set the environment variable for the decode instance.
  tests/e2e/test_ernie_03b_pd_splitwise_scheduler.py:249
• Inconsistent spacing in f-string expression. Line 249 has FD_API_PORT+1 without spaces around the + operator, while line 249's first URL uses proper spacing. For consistency, use FD_API_PORT + 1.