refactor: eliminate DisaggPDChunkedPrefillScheduler.

[weizhehuang0827]

Description

Background

After the scheduler policy refactor (#1992), DisaggPDChunkedPrefillScheduler became the last scheduler subclass with its own prepare_batch() that bypasses the SchedulerPolicy abstraction. Its core responsibility was:

1. Prevent hold-and-wait deadlock: In PD disagg mode, completed prefills must transfer KV to decode before releasing blocks. If multiple large requests each grab a partial chunk, none can complete → deadlock. The class enforced pd_prefill_footprint_fits to gate fresh request admission.
2. Prioritize in-flight over fresh: Using a "deferred re-queue" pattern to skip blocked fresh requests while still giving in-flight chunked requests their tokens.

Both concerns are now handled by the general PrefillFirstPolicy:

• Deadlock prevention: PrefillFirstPolicy schedules chunk_queue (in-flight) before prefill_queue (fresh). In-flight requests always get their chunks first → complete → release blocks → fresh can start. A new full-footprint admission gate in schedule_prefill_from_queue provides additional protection.
• Priority ordering: Separate chunk_queue / prefill_queue naturally eliminates the need for deferred re-queue — a blocked fresh request in prefill_queue never starves an in-flight request in chunk_queue.

Full-Footprint Admission Gate

Added inside schedule_prefill_from_queue, gating fresh requests only (kv_cache_tokens_num == 0):

if (batch_mode_.enable_chunked_prefill &&
    request->sequences()[0]->kv_state().kv_cache_tokens_num() == 0) {
  if (reserved_full_footprint + full_footprint > total_blocks) {
    blocks_exhausted = true;
    break;
  }
}
// After successful admission:
reserved_full_footprint += full_footprint;

• reserved_full_footprint is passed by reference from the policy layer, accumulating across both chunk_queue and prefill_queue scheduling passes.
• In DecodeFirstPolicy, it's initialized with max_used_blocks * 1.1 to reserve margin for decode token generation (reduces preemption frequency).
• In PrefillFirstPolicy, it starts at 0 (no decode on PD PREFILL instance).
• In-flight chunked requests (kv_cache_tokens_num > 0) skip this check — they are already admitted and must be allowed to continue.

Other Adaptations

• drain_request_queue: Skip expand_sequences for PD PREFILL instances (best_of_n expansion deferred to DECODE instance, preserving the optimization from the deleted class).
• resolve_batch_mode: PD PREFILL instances always use PrefillFirstPolicy (exclusive batch), regardless of enable_mix_batch setting.
• handle_unschedulable_head: Removed chunk_queue.empty() condition — if a stuck request in chunk_queue can never complete (footprint > total), it should also be error-reported.
• record_num_prefix_cache_tokens(): Already present in schedule_prefill_from_queue (line 322), preserving the prefix cache propagation fix from bugfix: propagate prefix cache usage across pd. #2011.

What's Deleted

Deleted	Replacement
DisaggPDChunkedPrefillScheduler class	DisaggPDScheduler → ContinuousScheduler::prepare_batch() → PrefillFirstPolicy
DisaggPDScheduler::prepare_batch() override	Inherited from ContinuousScheduler (drain handled by drain_request_queue)
pd_prefill_footprint_fits()	Full-footprint gate in schedule_prefill_from_queue
pd_prefill_remaining_blocks()	Not needed (scheduling order prevents deadlock)
pick_pd_chunk_budget()	compute_prefill_tokens()
alloc_chunk() / match_prefix_blocks()	allocate_for_prefill() / allocate_shared_blocks_for()
schedule_waiting_prefill() (deferred pattern)	schedule_prefill_from_queue (chunk_queue first, then prefill_queue)
DISAGG_PD_CHUNKED_PREFILL enum	Removed from SchedulerKind

Related Issues

Follow-up to #1992 (scheduler policy refactor)

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format.

Pre-commit Checks

• pre-commit run --all-files passed.

Self Review

• Self-reviewed per custom-code-style.md.
• Rebased onto latest main.

Build and Test Coverage

• Tests added (FullFootprintAdmissionGate, FullFootprintAdmitsBothWhenFits).
• NPU: unit tests (42 PASSED) + PD serving stress tests passed.

Reviewer Notes

Test Results

Unit tests: 42 PASSED (scheduler_test)

PD Stress Tests (Qwen3-8B, single card, etcd + xllm_service):

Case	Dataset	Requests	Concurrency	Result
PD chunked=true	ShareGPT	100	20	PASSED
PD chunked=true	Random 2000in/100out	50	15	PASSED
PD chunked=false	ShareGPT	100	20	PASSED
PD chunked=false	Random 2000in/100out	50	15	PASSED