shaun-z/experiment-queue

Experiment Queue

Orchestrate large batches of ML experiments on SSH remote GPU servers with proper state tracking, OOM retry, stale cleanup, and wave transitions.

When to Use This Skill

Use when /run-experiment is insufficient:

• ≥10 jobs that need batching across GPUs
• Multi-seed sweeps (e.g., 21 seeds × 12 cells)
• Wave transitions (run wave 1, wait, run wave 2, wait, run wave 3...)
• Teacher+student chains (train teacher then distill; auto-trigger student after teacher done)
• OOM-prone configs where you need to retry with different GPU or wait
• Mixed seed grids where failed cells need re-running

Do NOT use for:

• Single ad-hoc experiment (use /run-experiment)
• Modal/Vast.ai deployments (those have their own orchestration)
• Experiments that need manual inspection between runs

Why This Exists

Based on session audit (2026-04-16), the major wall-clock sinks in multi-seed grid experiments are:

1. Stale screens — python finishes, wandb uploads, screen hangs, next wave blocked
2. OOM on shared GPU — previous job's memory not yet released
3. Wave race — new wave launches before previous wave fully settles
4. Missing checkpoints — student launches before teacher saved
5. Parser duplication — rewriting multi-seed analysis python every batch

All of these are pure engineering friction that can be orchestrated.

Core Concepts

Job Manifest

A manifest lists jobs with explicit state:

project: dllm_distill
cwd: /home/rfyang/rfyang_code/dllm_experiments_torch
conda: dllm
# Optional: override conda hook path if conda is not at a standard location.
# Can be a bare path (wrapped automatically) or a full `eval "$(... shell.bash hook)"` string.
# Falls back to auto-detect of ~/anaconda3, ~/miniconda3, /opt/anaconda3, etc.,
# or the ARIS_CONDA_HOOK environment variable.
# conda_hook: /custom/path/to/conda
ssh: SJTUServer5
default_cmd: >
  python run_pc_distill_exp.py --backbone softmax --lam 0.5
  --K 500 --L 96 --W 16 --n_steps 30000 --batch_size 128 --lr 1e-4

preconditions:
  - type: checkpoint_exists
    path: checkpoints/transformer/pcc_softmax_L96_K500_N{N}_wikitext103.pt

gpus: [0, 1, 2, 3, 4, 5, 6, 7]
max_parallel: 8
gpu_free_threshold_mib: 500  # optional, default 500; raise for shared servers, lower for tight packing
oom_retry:
  delay: 120
  max_attempts: 3

jobs:
  - id: s200_N64_n50K
    args: {seed: 200, n_hidden: 64, n_train_subset: 50000, subset_seed: 2024}
  - id: s200_N128_n50K
    args: {seed: 200, n_hidden: 128, n_train_subset: 50000, subset_seed: 2024}
  # ... 14 more

Job State Machine

pending → running → completed
                 ↘ failed_oom → pending (after delay) [retry up to N]
                 ↘ failed_other → stuck (needs manual inspection)
stale_screen_detected → cleaned → pending

Wave Orchestration

A "wave" is a batch of jobs that fit available GPUs. Next wave only starts when:

1. All current-wave python processes have exited
2. No stale screens remain for current-wave tags
3. GPU memory has dropped below threshold (≤500 MiB)
4. Precondition checks pass for next-wave jobs

Workflow

Step 1: Parse Manifest / Build from Grid

Input can be:

• YAML manifest (explicit job list, recommended for complex cases)
• Grid spec (Cartesian product of param values, e.g., N=[64,128,256] × n=[50K,150K,500K,652K])
• Natural language description (Claude parses into manifest)

Save the built manifest to <project>/experiment_queue/<timestamp>/manifest.json for reproducibility.

Step 2: Pre-flight

• Check SSH connection works
• Check conda env exists on remote
• Check cwd exists on remote
• Check all preconditions (checkpoints, input files)
• Check GPU availability (at least max_parallel free GPUs)

If any precondition fails, show user which jobs are blocked and why.

Step 3: Launch Scheduler

Run tools/queue_manager.py (bundled with this skill) as a detached nohup process on the SSH host:

ssh <server> 'nohup python3 ~/.aris_queue/queue_manager.py \
  --manifest /tmp/manifest.json \
  --state /tmp/queue_state.json \
  --log /tmp/queue.log \
  > /tmp/queue_mgr.log 2>&1 &'

The scheduler:

• Reads manifest
• Loops: for each pending job, assign to free GPU, launch via screen
• Polls job status (every 60s)
• Detects stale screens (python exited but screen detached → kill)
• Detects OOM (CUDA OOM in log → mark failed_oom → retry after delay)
• Detects completion (expected output JSON/file exists) → mark completed
• Launches next wave when current wave settles
• Writes state to queue_state.json continuously

Step 4: Monitoring

User can check state anytime:

ssh <server> cat /tmp/queue_state.json | jq '.jobs | group_by(.status) | map({(.[0].status): length}) | add'

Or invoke /monitor-experiment which reads the state file.

Step 5: Post-completion

When all jobs in manifest.json are completed or stuck:

• Scheduler exits cleanly
• Write final summary to <project>/experiment_queue/<timestamp>/summary.md
• Invoke /analyze-results if analyze_on_complete: true

Grid Spec Syntax

Instead of writing 24 job entries manually:

grid:
  N: [64, 128, 256]
  n: [50000, 150000, 500000, 652000]
  seed: [42, 200, 201]
template:
  id: "s${seed}_N${N}_n${n}"
  args: {seed: ${seed}, n_hidden: ${N}, n_train_subset: ${n}}

Expands to 36 jobs automatically.

Wave Chaining

For sequential phases (teacher → student):

phases:
  - name: train_teachers
    grid:
      N: [384, 512]
    template:
      cmd: python run_pc_exp.py --direction c --backbone softmax --n_hidden ${N} ...
      output_check: checkpoints/transformer/pcc_softmax_L96_K500_N${N}_wikitext103.pt
  
  - name: distill_students
    depends_on: train_teachers
    grid:
      N: [384, 512]
      seed: [42, 200, 201]
    template:
      cmd: python run_pc_distill_exp.py --n_hidden ${N} --seed ${seed} ...
      output_check: figures/pcdistill_sw_N${N}_*_seed${seed}.json

Scheduler enforces depends_on: distill_students jobs stay pending until all train_teachers jobs are completed.

OOM Handling

Detect OOM from stdout:

torch\.OutOfMemoryError: CUDA out of memory

On detection:

1. Mark job failed_oom
2. Kill screen
3. Wait oom_retry.delay seconds
4. Check if current GPU is free; if not, try another free GPU
5. Requeue as pending
6. Max oom_retry.max_attempts before marking stuck

Stale Screen Detection

Every 60s, for each running screen:

1. Check screen exists (screen -ls)
2. Check python PID still running (ps -p)
3. If screen exists but python exited:
   • If expected output file exists → mark completed, kill stale screen
   • If no output file → mark failed_other, kill screen

Resume-on-restart

If scheduler crashes / is killed:

1. Read queue_state.json
2. For each running job: check screen; if still alive, keep; if not, re-evaluate state
3. For each pending: continue normally
4. Idempotent: safe to restart scheduler without losing state

Output: Summary Report

# Experiment Queue Summary

**Project**: dllm_distill
**Started**: 2026-04-16 11:36:29
**Completed**: 2026-04-16 18:02:14
**Total wall-clock**: 6h 25m
**Jobs**: 40 completed, 2 OOM-retried then completed, 0 stuck

## Phases
| Phase | Jobs | Success | OOM retries | Duration |
| --- | --- | --- | --- | --- |
| train_teachers | 2 | 2 | 0 | 58m |
| distill_students | 24 | 24 | 2 | 4h 02m |
| multi_seed_validation | 16 | 16 | 0 | 1h 25m |

## Results Files
- 42 JSON files in `figures/pcdistill_sw_*.json`

## Next Steps
- Run `/analyze-results` on output JSONs
- Figures auto-regen via `artifact-sync` (if configured)

Comparison with /run-experiment

| Feature | /run-experiment | experiment-queue | | --- | --- | --- | | Single-shot experiment | ✅ | ✅ (overkill) | | Multi-GPU parallel | Basic | Proper scheduling | | Wave transitions | Manual | Automatic | | OOM retry | Manual | Automatic | | Stale screen cleanup | Manual | Automatic | | Teacher→student chain | Manual | Built-in | | State persistence | No | Yes (JSON) | | Resume on crash | No | Yes | | Grid expansion | Manual | Declarative |

Rule: Use /run-experiment for ≤5 jobs. Use experiment-queue for ≥10 jobs or anything with phases.

Key Rules

• Never overlap screens on the same GPU — always wait for memory.used < 500 MiB before launching new job
• Always write state to disk — every state change flushed to queue_state.json
• Idempotent scheduler — safe to restart; picks up from state file
• Expected-output-based completion — don't trust screen state alone; verify output file exists
• Bounded retry — max N OOM retries, then mark stuck and alert
• Dependencies enforced at launch — never launch student before teacher checkpoint exists

Known Failure Modes

• SSH connection drop during scheduling: scheduler keeps running on remote (nohup), just reconnect and check
• GPU reservation by another user: scheduler waits, does not pre-empt
• Disk full on remote: scheduler detects write failure, marks all pending stuck, alerts

Example Session

User: "跑 T5+T6 全部实验：T5 = N∈{80,192} × n 4 values × seed {200,201}, T6 = N∈{384,512} × n 4 values × seed {42,200,201}; T6 需要先 train teacher"

Claude invokes /experiment-queue:

1. Parses description into 2-phase manifest
2. Phase 1: T5 (16 jobs, no teacher dependency) + T6 teacher training (2 jobs)
3. Phase 2: T6 distillation (24 jobs, depends on teachers)
4. Deploys scheduler via nohup
5. Reports: "Scheduler PID 93534, total 42 jobs, estimated 6-7h wall-clock"

Then user can check anytime or wait for summary report.

See Also

• /run-experiment — single experiment deployment
• /monitor-experiment — check progress (now reads from queue_state.json)
• /analyze-results — post-hoc analysis
• tools/queue_manager.py (bundled) — the scheduler implementation
• tools/build_manifest.py (bundled) — build manifest from grid spec

Rationale / Source

Identified via 2026-04-16 post-mortem analysis (Codex GPT-5.4 xhigh) of a 1.5-day multi-seed paper experiment session:

• Wall-clock sink: stale screens, OOM, wave transitions, manual parser
• Token sink: re-writing orchestration code each session
• Cognitive sink: tracking which cells succeeded, which failed, which to retry

This skill targets the wall-clock sink specifically; see artifact-sync and paper-fix-auto-apply for the other two.