Tianye/4gpu 2ppl fix on open pipelines

[TianyeGGBond]

Between rollouts, actor_train preempts shared infer GPUs, causing both pipelines' actor_infer DP workers to shrink to set().

The gap-ratio planner needs a fresh demand signal to re-wake engines, but that signal only arrived from begin_progress_batch inside the rollout function — after sample dispatch had already started.

Race result: whichever pipeline fires begin_progress_batch first wins all DP workers; the other pipeline hangs.

Two sub-bugs compounded it:

• pending_bucket_gen is a per-cycle snapshot that vanishes once consumed, so the planner sees no demand during the rollout boundary gap.
• MilesRLixHooks was never wired at init, so every begin_progress_batch hit NoOpRLixHooks; the scheduler never received rollout demand at all.

sequenceDiagram
    participant P1 as Pipeline 1
    participant P2 as Pipeline 2
    participant Sched as Scheduler (gap-ratio)

    Note over P1,Sched: _after_training: actor_train preempts infer GPUs
    Note over Sched: active_dp_ranks = ∅ for both<br/>pending_bucket_gen consumed → planner skips both

    P1->>Sched: begin_progress_batch()  ← first to fire
    Sched->>Sched: wakes P1 engines, allocates all DP workers ✓

    P2->>Sched: begin_progress_batch()  ← too late
    Sched-->>P2: ❌ P1 holds all DP workers
    Note over P2: "No progress for 30.0s. Queue: 0, Collected: 0/N"

Loading

---

Fix

Pre-signal scheduler demand before the rollout function starts, using request_gpus instead of report_progress.

This makes the call write to the new durable rollout_open_pipelines registry and block until at least one DP worker is awake.

sequenceDiagram
    participant P1 as Pipeline 1
    participant P2 as Pipeline 2
    participant Sched as Scheduler

    Note over P1,Sched: _after_training: actor_train preempts infer GPUs
    Note over Sched: active_dp_ranks = ∅ for both

    P1->>Sched: signal_rollout_demand() → request_gpus(GENERATION, step_target=N)
    P2->>Sched: signal_rollout_demand() → request_gpus(GENERATION, step_target=N)
    Note over Sched: rollout_open_pipelines = {P1: N, P2: N} ← durable<br/>gap-ratio sees both → allocates DP workers for both

    Sched-->>P1: unblocks ✓ (engines awake)
    Sched-->>P2: unblocks ✓ (engines awake)

    P1->>P1: rollout fn dispatches first sample ✓
    P2->>P2: rollout fn dispatches first sample ✓

Loading

---

Changes

Commit	What
dfd53f3	Replace pending_bucket_gen with durable rollout_open_pipelines (pipeline_id → step_target) in SchedulerState; planner reads it for the full rollout lifecycle.
f2ac106	Wire MilesRLixHooks in _init_phase_b_infer; add no-restart tombstones to Orchestrator / SchedulerActor.
173dbcd	Add MilesPipeline.signal_rollout_demand() to pre-signal demand before rollout dispatch.
8436e95	Fix signal_rollout_demand to call request_gpus instead of report_progress.
e6276b9	Add tests: static guard for no runtime Orchestrator dependency in scheduler, plus chaos test where orchestrator kill leaves scheduler reachable.

---

Requires companion miles PR

miles@tianye/4gpu-2ppl-fix

This companion PR wires:

• rlix_train_loop.signal_demand hook
• run_miles_dual call-site

---

Test

Unit tests

• test_orchestrator_death_is_benign.py
• test_gap_ratio.py
• test_scheduler_apply_plan_invariants.py

E2E test

run_4ppl_full_overlap_n10.sh --num-rollout 5

Configuration:

• P1 train = [0, 1]
• P1 infer = [0, 1, 2, 3]
• P2 train = [2, 3]
• P2 infer = [0, 1, 2, 3]

Result:

• Run 28: all 5 rollouts clean, EXIT=0 ✓