File size: 6,266 Bytes
0784852 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 | """Run actual isolated CUDA preflights and read-only ordinary endpoint checks."""
from __future__ import annotations
import argparse
import hashlib
import json
import subprocess
from pathlib import Path
from execute_learning_use import claim_gpu
from run_memory_interface_next import now, write
def validate(config_path):
config_path = Path(config_path).resolve()
config = json.loads(config_path.read_text())
project = Path(__file__).resolve().parents[1]
artifact = project / "docs/development-artifacts" / config["batch"]
source = Path(config.get("source_root", project))
jobs = [
(spec["name"], index, gpu, False)
for index, (spec, gpu) in enumerate(zip(config["specs"], config["gpus"], strict=True))
]
jobs.append(("reused-ordinary", None, 4, True))
locks, workers = [], []
try:
for name, index, gpu, reused in jobs:
lock = claim_gpu(gpu)
if lock is None:
raise RuntimeError(f"GPU {gpu} is occupied; no other experiment is preempted")
locks.append(lock)
out = artifact / "preflight" / name
out.mkdir(parents=True, exist_ok=False)
container = "lm-moe-preflight-" + hashlib.sha256(str(out).encode()).hexdigest()[:16]
command = [
"docker",
"--context",
"lm-memory",
"run",
"--name",
container,
"--label",
"project=llm-memory-editability",
"--label",
"batch=" + config["batch"],
"--network=none",
"--gpus",
f"device={gpu}",
"--cpus",
"4",
"--memory",
"32g",
"--read-only",
"--tmpfs",
"/tmp:rw,size=2g",
"--shm-size=1g",
"--mount",
f"type=bind,src={project},dst={project},readonly",
"--mount",
f"type=bind,src={out},dst={out}",
"--workdir",
str(project),
"--env",
"MEMORY_INTERFACE_CONTAINER=1",
"--env",
"PYTHONDONTWRITEBYTECODE=1",
"--env",
"OMP_NUM_THREADS=1",
"--env",
"PYTHONPATH=" + str(source / "src"),
config["runtime"]["image"],
config["runtime"]["python"],
"-u",
"-m",
"llm_memory_editability.moe_sequential",
"--config",
str(config_path),
"--out",
str(out),
"--device",
"cuda:0",
]
if reused:
command.append("--verify-reused")
else:
command += ["--preflight", "--run", config["specs"][index]["name"]]
write(out / "command.json", {"command": command, "physical_gpu": gpu})
log = (out / "output.log").open("w")
process = subprocess.Popen(command, stdout=log, stderr=subprocess.STDOUT)
workers.append((name, out, process, log, reused))
results = []
for name, out, process, log, reused in workers:
returncode = process.wait()
log.close()
path = out / ("reuse-validation.json" if reused else "preflight.json")
result = json.loads(path.read_text()) if path.exists() else {}
passed = returncode == 0 and result.get("passed") is True
results.append(
{"name": name, "passed": passed, "returncode": returncode, "evidence": str(path)}
)
# Check real-sized ordinary, expert and wide parameter counts and common tensors.
ledgers = {}
for entry in config["reused_runs"]:
ledger = artifact / "preflight/reused-ordinary" / entry["name"] / "architecture.json"
ledgers.setdefault("D4", json.loads(ledger.read_text()))
for spec in config["specs"]:
ledger = artifact / "preflight" / spec["name"] / "architecture.json"
if ledger.exists():
ledgers.setdefault(spec["architecture"], json.loads(ledger.read_text()))
counts = {
arm: {
key: ledger[key]
for key in ("total_parameters", "nominal_parameters_selected_per_token")
}
for arm, ledger in ledgers.items()
}
matching = False
if set(ledgers) == {"D4", "M4", "W4"}:
m, w, d = (ledgers[arm] for arm in ("M4", "W4", "D4"))
matching = (
abs(m["total_parameters"] - w["total_parameters"]) < 0.001 * m["total_parameters"]
)
matching &= (
abs(m["nominal_parameters_selected_per_token"] - d["total_parameters"])
< 0.001 * d["total_parameters"]
)
dense = {
key: value
for key, value in d["initialization"]["shared_tensor_sha256"].items()
if ".mlp." not in key
}
matching &= dense == m["initialization"]["shared_tensor_sha256"]
matching &= dense == w["initialization"]["shared_tensor_sha256"]
summary = {
"passed": matching and all(row["passed"] for row in results),
"jobs": results,
"real_size_parameter_ledger": counts,
"paired_common_tensors_and_parameter_matching": matching,
"created_utc": now(),
}
write(artifact / "preflight-summary.json", summary)
print(json.dumps(summary), flush=True)
if not summary["passed"]:
raise SystemExit(1)
finally:
# Running isolated preflights are preserved if orchestration fails.
for _, _, process, log, _ in workers:
if process.poll() is not None:
log.close()
for lock in locks:
lock.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--config", type=Path, required=True)
validate(parser.parse_args().config)
|