Seer Oracle Replay
Reproduce the Project Seer teacher-only resource study, then extend the scheduling experiments without the original GPU cluster. This repository contains the analysis code and handoff; namezz/seer-agent-traces contains immutable inputs and reference results.
Quick start
Python 3.10 or newer; CPU only. No CUDA, Docker, Slurm, model download, or new inference is needed for offline replay. The published run takes roughly 1–2 minutes on the original CPU host; runtime varies by machine.
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
# Public dataset; no HF login is required.
# Replace COMMIT with the full revision in DATA_RELEASE.json:
seer-download --revision COMMIT --dest data
seer-replay reproduce --data-dir data/replay --output-dir runs/reproduction
python tests/check_reproduction.py runs/reproduction
The replay writes RESULT.json, independently checked schedules and VALIDATION.json, two workload-specific PNG/PDF/SVG plots, and joint_schedules.csv. The same color within a plot means the same complete schedule. The reproduction test compares all 28 published resource/time/bound/reduction values.
To examine full model/tool logs, add --with-raw to the download command. The compact layer is 3.79 MiB compressed (approximately 11 MB unpacked); the original compressed teacher archives are approximately 10.34 GB. See docs/DATA.md for contents and exclusions.
What was evaluated?
| Workload | Teacher | Scope | Included |
|---|---|---|---|
| Terminal-Bench 2.1 | TMax-27B, 27B | 89 tasks × 3 seeds | 253 normal trials; 14 execution exceptions separate |
| SWE-bench Verified | DeepSWE-Preview, 32B | 500 tasks × 4 attempts | 1,998 generations; 2 infrastructure failures separate |
DeepSWE includes generation only, excluding downstream official grading. TMax includes the recorded trial lifecycle including verifier/teardown. Model mistakes and ordinary step/token limits remain included. Do not directly compare the two lifecycle definitions as if they were identical.
What the numbers mean
These are conditional offline replay estimates of reservations, not measured utilization or deployed speedups. Per-task durations and internal model-request intervals remain fixed. CPU/RAM are sandbox quotas; waiting sandboxes are assumed to pause/offload with zero cost. GPU-hours count an 8-GPU service while any model request is pending. Storage covers only shared compressed OCI image blobs, not total disk or RAM-offload storage.
Each objective is optimized separately, subject to completion no later than FIFO and unchanged resource caps. CPU/RAM attain analytical minima in this model. GPU/cache results are the best feasible schedules found, not globally proven minima. See docs/METHODS.md and docs/HANDOFF.md before extending the results.
Results retained for handoff
| Workload / objective | Resource-hours | Completion (min) | FIFO deadline (min) |
|---|---|---|---|
| TMax CPU | 19.8786 core-h | 110.3300 | 132.8094 |
| TMax RAM | 54.0641 GiB-h | 110.3300 | 132.8094 |
| TMax GPU | 14.3611 GPU-h | 129.2053 | 132.8094 |
| DeepSWE CPU | 148.0393 core-h | 1,311.4622 | 1,352.9339 |
| DeepSWE RAM | 592.1572 GiB-h | 1,311.4622 | 1,352.9339 |
| DeepSWE GPU | 173.2963 GPU-h | 1,311.8427 | 1,352.9339 |
| DeepSWE image cache | 180.1541 GiB-h | 1,313.4180 | 1,352.9339 |
CPU/RAM minima coexist with every selected oracle schedule. TMax's GPU schedule achieves all three measured minima together. DeepSWE's GPU and cache best-found values belong to different schedules; read joint_schedules.csv for simultaneously attained values.
Repository map
src/seer_replay/search.py: admission scheduling, objective accounting, seeded order search.refine_storage.py: additional shared-image task grouping search.validate.py: independent scalar interval-union and capacity/deadline validation.plot_schedules.py: complete-schedule comparisons, independent workload axes.download.py: pinned HF download and SHA256 verification.provenance/original-analysis/: original scripts, retained for audit; not portable entrypoints.docs/: workload configuration, schema, assumptions, next experiments, team handoff.
No credentials, unrelated Codex conversation, model weights, Docker images, or Python environments are part of this project. No new benchmark/training jobs are started by these commands.