File size: 4,239 Bytes
20377b3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 | #!/usr/bin/env python3
"""Run a separately registered looped-Transformer comparison."""
import argparse
import fcntl
import json
import os
import platform
import shutil
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "src"))
def main():
from llm_memory_editability.grok_depth import source_hash, utc, write_json
from llm_memory_editability.grok_loop_data import build_world
from llm_memory_editability.grok_loop_train import run
from llm_memory_editability.grok_multihop_data import audit_world
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("run_id")
parser.add_argument("--config", default="configs/grok-loop-development-v1.json")
parser.add_argument("--device", default="cuda:0")
parser.add_argument("--resume", action="store_true")
parser.add_argument("--data-only", action="store_true")
args = parser.parse_args()
cfg = json.loads((ROOT / args.config).read_text())
spec = {**cfg["base"], **cfg["runs"][args.run_id]}
if sorted(set(spec["nodes"])) != spec["nodes"] or spec["nodes"][-1] != spec["steps"]:
raise ValueError("Registered evaluation nodes must be sorted and include the endpoint")
out = ROOT / cfg["output_root"] / spec["phase"] / args.run_id
files = [
ROOT / p
for p in [
args.config,
"src/llm_memory_editability/grok_multihop.py",
"src/llm_memory_editability/grok_multihop_data.py",
"src/llm_memory_editability/grok_depth.py",
"src/llm_memory_editability/grok_depth_data.py",
"src/llm_memory_editability/bios_model.py",
"scripts/run_grok_loop.py",
"src/llm_memory_editability/grok_loop_model.py",
"src/llm_memory_editability/grok_loop_train.py",
"src/llm_memory_editability/grok_loop_data.py",
]
]
lock = json.loads((ROOT / cfg["source_lock"]).read_text())
for relative, expected in lock["files"].items():
path = ROOT / relative
if source_hash([path])[str(path)] != expected:
raise ValueError(f"Locked execution source changed: {relative}")
world = build_world(spec)
if args.data_only:
print(json.dumps(audit_world(world), indent=2))
return
if args.resume:
meta = json.loads((out / "metadata.json").read_text())
if meta["spec"] != spec:
raise ValueError("Resume spec differs from original metadata")
for path in files[1:]:
if source_hash([path])[str(path)] != meta["files"][str(path)]:
raise ValueError(f"Resume source changed: {path}")
else:
out.mkdir(parents=True, exist_ok=False)
for path in files:
dest = out / "source" / path.relative_to(ROOT)
dest.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(path, dest)
write_json(
out / "metadata.json",
{
"started_utc": utc(),
"spec": spec,
"files": source_hash(files),
"pid": os.getpid(),
"visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
"argv": sys.argv,
"python": platform.python_version(),
},
)
lock_handle = (out / "run.lock").open("a")
fcntl.flock(lock_handle, fcntl.LOCK_EX | fcntl.LOCK_NB)
write_json(out / "data-audit.json", audit_world(world))
write_json(out / "world-metadata.json", world["metadata"])
import numpy as np
import torch
write_json(
out / "environment.json",
{
"torch": torch.__version__,
"numpy": np.__version__,
"cuda": torch.version.cuda,
"device": str(args.device),
"gpu": torch.cuda.get_device_name(torch.device(args.device)),
"tf32": True,
"precision": "FP32 parameters, forward and optimizer; TF32 matmuls",
},
)
np.savez_compressed(
out / "world.npz", **{k: v for k, v in world.items() if isinstance(v, np.ndarray)}
)
run(spec, world, out, args.device, args.resume)
if __name__ == "__main__":
main()
|