104 lines
3.1 KiB
TOML
104 lines
3.1 KiB
TOML
[run]
|
|
name = "aggressive_oom_sweep_probe_node"
|
|
timeout_minutes = 1440
|
|
local_artifact_dir = "artifacts/remote_runs"
|
|
max_attempts = 1
|
|
artifact_sync_interval_seconds = 60
|
|
|
|
[provider]
|
|
kind = "vastai"
|
|
disk_gb = 192
|
|
max_price_per_hour = 0.80
|
|
image = "vastai/base:0.0.2"
|
|
|
|
[provider.gpu]
|
|
name = "RTX 4090"
|
|
count = 1
|
|
min_vram_gb = 20
|
|
|
|
[selection]
|
|
min_reliability = 0.95
|
|
min_down_mbps = 100
|
|
min_up_mbps = 25
|
|
require_verified = true
|
|
blocked_geos = ["CN"]
|
|
blacklist_hosts = [59017, 1647, 92578, 1276, 75481, 1256, 85323, 34031]
|
|
drop_cheap_frac = 0.30
|
|
image_size_gb = 5.0
|
|
base_url = "https://cloud.vast.ai"
|
|
|
|
[workspace]
|
|
workdir = "."
|
|
exclude = [
|
|
"/artifacts/remote_runs",
|
|
"/artifacts/runs",
|
|
"/artifacts/data_cache",
|
|
"/artifacts/model_sanity",
|
|
"/artifacts/current_run",
|
|
"/artifacts/public_airfrans",
|
|
"/artifacts/notes",
|
|
"/artifacts/preflight_20260725",
|
|
"/artifacts/remote_state_20260725",
|
|
"/artifacts/sweep_20260725",
|
|
"/artifacts/model_sanity_decisions",
|
|
"/artifacts/model_sanity_verify",
|
|
"/artifacts/*.json",
|
|
"/artifacts/*.yaml",
|
|
"/artifacts/*.md",
|
|
"/data/raw",
|
|
"/data/processed",
|
|
"/.venv",
|
|
"/.git",
|
|
"/.airfrans_hf_resume",
|
|
"/sky_logs",
|
|
"/notebooks",
|
|
"__pycache__",
|
|
"*.pyc",
|
|
]
|
|
|
|
[bootstrap]
|
|
command = """
|
|
uv sync --no-dev
|
|
uv run --no-dev python -c "import torch; ok=torch.cuda.is_available(); print('torch_cuda_available=' + str(ok)); print(torch.cuda.get_device_name(0) if ok else 'no_cuda_device'); assert ok, 'torch CUDA unavailable'"
|
|
uv run --no-dev python -c "import huggingface_hub, wandb; print('hf_wandb_import_ok')"
|
|
"""
|
|
|
|
[data]
|
|
validation_command = """
|
|
uv run --no-dev python -c "from airfrans_frontier.sweep import read_jobs; jobs=read_jobs('artifacts/aggressive_oom_sweep/jobs.jsonl'); assert len(jobs) == 34; assert any(j['status'] == 'pending' for j in jobs); print('pending_jobs=' + str(sum(1 for j in jobs if j['status'] == 'pending')))"
|
|
uv run --no-dev python -c "from airfrans_frontier.training.config import load_training_config; c=load_training_config('artifacts/aggressive_oom_sweep/configs/100m_mlp_raw.toml'); assert c.data.source == 'huggingface'; assert c.huggingface.enabled; print('probe_config=' + c.run.name + ' source=' + c.data.source)"
|
|
"""
|
|
|
|
[job]
|
|
command = """
|
|
uv run --no-dev python scripts/aggressive_oom_node_wrapper.py --jobs artifacts/aggressive_oom_sweep/jobs.jsonl --node "$AIRFRANS_REMOTE_RUN_ID" --artifact-dir artifacts/current_run --utilization artifacts/aggressive_oom_sweep/utilization.jsonl --max-jobs 1 --sample-interval-seconds 30 --require-success
|
|
"""
|
|
artifact_dir = "artifacts/current_run"
|
|
heartbeat_file = "artifacts/current_run/heartbeat.json"
|
|
metrics_file = "artifacts/current_run/metrics.jsonl"
|
|
|
|
[artifacts]
|
|
mode = "object_store_upload"
|
|
required = [
|
|
"config.toml",
|
|
"job_manifest.json",
|
|
"metrics.jsonl",
|
|
"latest_metrics.json",
|
|
"heartbeat.json",
|
|
"checkpoint_latest.pt",
|
|
"checkpoint_best.pt",
|
|
"checkpoint_final.pt",
|
|
"final_metrics.json",
|
|
"run_manifest.json",
|
|
"environment_manifest.json",
|
|
"utilization.jsonl",
|
|
"node_summary.json",
|
|
"sweep_collect.json",
|
|
"sweep_jobs.jsonl",
|
|
"sweep_attempts.jsonl",
|
|
"sweep_utilization.jsonl",
|
|
]
|
|
|
|
[cleanup]
|
|
on_success = "sky_down"
|
|
on_failure = "collect_then_keep"
|