Worker decouple — Phase 2 / Slice 9: DecodeScheduler

[Andrewxu313]

Phase 2 of the worker decouple initiative. Final slice.

Scope

Extract the decode loop (worker:5592–6427 + decode-side of PAGE BOUNDARY, ~1500 LOC) into batchgen/worker/decode.py.

Design

class ModelExecutorBackend(Protocol):  # reused from Slice 6
    def forward_decode(self, batch) -> DecodeOutputs: ...

@dataclass(frozen=True)
class DecodeStepRequest:
    active_uuids: tuple[str, ...]
    cache_seqlens: torch.Tensor
    position_ids: torch.Tensor
    cuda_graph_decision: GraphDecision  # from Glm5CudaGraphAdapter (Phase A/B/C)

class DecodeScheduler:
    def __init__(self, *, executor: ModelExecutorBackend, sync: SyncCoordinator): ...
    def step(self, req: DecodeStepRequest) -> DecodeStepResult: ...

Composes with the cuda-graph adapter that already exists (Phase A/B/C Glm5CudaGraphAdapter). The handler's step calls adapter.eligibility(batch_state) → adapter.prepare_replay_inputs(...) → manager.replay(...) exactly as the current worker does.

3-PR plan

• PR-9.1 Port.
• PR-9.2 Dual-path gate + compare-mode + H20 L2-2048 + L2-4096 back-to-back (critical: this is the hottest path).
• PR-9.3 Cleanup. After this PR the worker is ~6,000 LOC.

Pre-merge checklist (from scheduler-split Step 9 fix-up commits)

• All self.model refs translated to backend-injected handle.
• DualKVCacheCoordinator import resolved.
• Model synced back to worker after configure_decoding.
• parallel_manager access via Protocol, not hasattr(self, 'parallel_manager').
• _put_sequences_on_hold routed through HostKVRebalancer (Slice 7).

Design source reference

origin/tairan/scheduler-split:batchgen/worker/decode.py (878 LOC).
origin/tairan/worker-reextract:batchgen/worker/decode/ (decomposed into 9 sub-files) — out of scope here; single-file form for now.

Design: /Users/andrew/.claude/plans/the-task-for-this-declarative-melody.md (POIS-local). Builds on the Phase A/B/C cuda-graph contract pattern (PR #162/#163/#164, merged 2026-05-27).