File size: 18,925 Bytes
a7a9170 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | """Offline MPS optimization/export only. Go supplies all features and weights."""
import argparse
import copy
import json
import math
from pathlib import Path
import resource
import subprocess
import time
import numpy as np
import torch
import train_pilot_v2 as core
ARMS = ("uniform-initial", "set-initial", "set-feedback")
TEMPERATURES = (.5, 1., 2., 4.)
PROTOCOL = "docs/own-three-choice-completeness-preregistration-20261002.md"
PROTOCOL_SHA = "5d840f7b3379339d684b85b7896a895538bd1a030395a096af3e52c5fc781cd3"
STATES_SHA = "38930441c45ec03336a66c670028208e525131f385699140d793da96cbf4028b"
AUDIT_SHA = "dc1bc90a5468b2f45a4e32501f0814a73ee2be7de56481c1edea3b7d414bfc6a"
INITIAL_SEED, SHUFFLE_SEED = 20261031, 20261032
RAW_CAP = 768 << 20
def read_json(path):
return json.loads(path.read_text(), object_pairs_hook=core.unique_pairs)
def retained_bytes(root):
total = 0
for phase in (root / "runs").glob("own-three-*"):
core.require(phase.is_dir() and not phase.is_symlink(), "regular phase directory")
for path in phase.rglob("*"):
core.require(not path.is_symlink(), "no raw evidence symlinks")
if path.is_file():
total += path.stat().st_size
core.require(total <= RAW_CAP, "whole-study cap; retain prefix and stop")
return total
def save(path, value, root):
raw = (json.dumps(value, ensure_ascii=False, indent=2, sort_keys=True, allow_nan=False) + "\n").encode()
core.require(len(raw) <= 1 << 20 and retained_bytes(root) + len(raw) <= RAW_CAP,
"bounded metadata and whole-study raw cap")
core.require(not path.exists(), "immutable fresh optimizer output")
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("xb") as out:
out.write(raw)
def save_jsonl(path, rows, root):
core.require(not path.exists(), "fresh bounded parity journal")
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("xb") as out:
for row in rows:
raw = (json.dumps(row, ensure_ascii=False, sort_keys=True, allow_nan=False) + "\n").encode()
core.require(len(raw) <= 1 << 20 and retained_bytes(root) + len(raw) <= RAW_CAP,
"bounded parity line and whole-study raw cap")
out.write(raw)
out.flush()
def record_update(path, value, root):
raw = (json.dumps(value, sort_keys=True, allow_nan=False) + "\n").encode()
core.require(len(raw) <= 1 << 20 and retained_bytes(root) + len(raw) <= RAW_CAP-(1 << 20),
"reserve a failure receipt and preserve actual update prefix")
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("xb" if value["actual_stage_updates"] == 1 else "ab") as out:
out.write(raw)
def load_inputs(directory, audit_path):
manifest_raw = (directory / "manifest.json").read_bytes()
manifest, audit = read_json(directory / "manifest.json"), read_json(audit_path)
core.require(manifest["schema"] == "gooo/own-three-training-inputs/v1" and
manifest["student_state_rows"] == 10739 and manifest["feature_dim"] == 768 and
manifest["label_count"] == 8 and manifest["planned_optimizer_steps"] == 4800 and
audit["status"] == "PASS" and audit["manifest_sha256"] == core.sha(manifest_raw) and
audit["student_states_sha256"] == STATES_SHA and audit["teacher_audit_sha256"] == AUDIT_SHA and
audit["feature_float32_values_recomputed"] == 10739*768 and
audit["model_predictions"] == audit["optimizer_updates"] == audit["native_calls"] == 0,
"independent Go replay of all source-bound features required")
for name, pin in manifest["files"].items():
raw = (directory / name).read_bytes()
core.require(len(raw) == pin["bytes"] and core.sha(raw) == pin["sha256"], "exact Go optimizer input pin")
rows = [json.loads(line, object_pairs_hook=core.unique_pairs)
for line in (directory / "rows.jsonl").read_bytes().splitlines()]
core.require(len(rows) == 10739 and all(r["feature_row_index"] == i for i, r in enumerate(rows)), "complete fixed feature row order")
x = np.fromfile(directory / "features-f32le.bin", dtype="<f4").reshape(10739, 768)
core.require(np.isfinite(x).all(), "finite Go features")
return manifest, rows, x, read_json(directory / "parity-inputs.json")["rows"]
def state_from_blob(path):
array = np.fromfile(path, dtype="<f4")
core.require(array.shape == (18656,) and np.isfinite(array).all(), "fresh Go FP32 initializer")
at, result = 0, {}
for key, shape in (("first.weight", (24, 768)), ("first.bias", (24,)),
("last.weight", (8, 24)), ("last.bias", (8,))):
n = math.prod(shape)
result[key] = torch.from_numpy(array[at:at+n].copy().reshape(shape))
at += n
return result
def state_sha(state):
return core.sha(b"".join(state[key].detach().cpu().numpy().astype("<f4").tobytes()
for key in ("first.weight", "first.bias", "last.weight", "last.bias")))
def arrays(rows, x, split, arm):
field = "feedback_arm_row_weight" if arm == "set-feedback" else "initial_arm_row_weight"
selected = [i for i, r in enumerate(rows) if r["split"] == split and r[field] > 0]
groups, indices = {}, []
weights = np.asarray([rows[i][field] for i in selected], dtype=np.float32)
target = np.asarray([rows[i]["finite_soft_targets"] for i in selected], dtype=np.float32)
core.require(target.shape == (len(selected), 8) and np.isfinite(target).all() and
(target >= 0).all() and np.max(np.abs(target.sum(axis=1)-1)) < 1e-6, "full finite passing-mask targets")
for local, original in enumerate(selected):
groups.setdefault(rows[original]["program_contract_group"], []).append(local)
core.require(split == "train" or rows[original]["phase"] == "initial", "no holdout feedback optimization")
for group in sorted(groups):
ix = np.asarray(groups[group], dtype=np.int64)
core.require(abs(float(weights[ix].sum())-1) < 1e-6, "equal function group weight")
indices.append(ix)
expected = 1024 if split == "train" else 256
core.require(len(indices) == expected, "fixed disjoint train/calibration groups")
return x[selected], target, weights, indices
def row_loss(logits, target, uniform):
if uniform:
return -(target * torch.log_softmax(logits, dim=-1)).sum(dim=-1)
return torch.logsumexp(logits, dim=-1) - torch.logsumexp(logits.masked_fill(target <= 0, -torch.inf), dim=-1)
def passing_nll(logits, targets, temperature):
# Stable float64 CPU logsumexp is calibration arithmetic, not optimization.
z = logits.astype(np.float64) / temperature
z -= z.max(axis=1, keepdims=True)
all_log = np.log(np.exp(z).sum(axis=1))
passing = np.where(targets > 0, z, -np.inf)
maximum = passing.max(axis=1)
pass_log = maximum + np.log(np.exp(passing - maximum[:, None]).sum(axis=1))
return all_log-pass_log
def fit(initial_state, data, arm, qat, output, root):
torch.manual_seed(INITIAL_SEED)
model = core.Model(qat=qat)
model.load_state_dict(initial_state)
input_sha = state_sha(model.state_dict())
model.to("mps")
optimizer = torch.optim.AdamW(model.parameters(), lr=.001, weight_decay=.01)
train = [torch.from_numpy(a).to("mps") for a in data["train"][:3]]
cal_x = torch.from_numpy(data["calibration"][0]).to("mps")
cal_y, cal_w = data["calibration"][1:3]
group_rows = data["train"][3]
rng = np.random.default_rng(SHUFFLE_SEED)
best, selected = None, (math.inf, 0, 0.)
history, updates, allocated, driver = [], 0, 0, 0
started, before = time.perf_counter(), resource.getrusage(resource.RUSAGE_SELF)
for epoch in range(100):
order, losses = rng.permutation(len(group_rows)), []
for begin in range(0, len(order), 128):
groups = order[begin:begin+128]
ix = torch.from_numpy(np.concatenate([group_rows[int(i)] for i in groups])).to("mps")
optimizer.zero_grad(set_to_none=True)
loss = (row_loss(model(train[0][ix]), train[1][ix], arm == "uniform-initial") * train[2][ix]).sum() / len(groups)
core.require(bool(torch.isfinite(loss).cpu()), "finite MPS objective")
loss.backward()
optimizer.step()
updates += 1
losses.append(float(loss.detach().cpu()))
record_update(output / ("qat" if qat else "fp32") / "optimizer-updates.jsonl",
{"arm": arm, "stage": "qat" if qat else "fp32", "epoch": epoch+1,
"batch": begin//128+1, "function_groups": len(groups),
"actual_stage_updates": updates, "objective": losses[-1]}, root)
with torch.no_grad():
logits = model(cal_x).cpu().numpy()
candidates = [(float((passing_nll(logits, cal_y, t)*cal_w).sum())/256, epoch+1, t)
for t in TEMPERATURES]
choice = min(candidates)
if choice < selected:
selected = choice
best = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
allocated = max(allocated, torch.mps.current_allocated_memory())
driver = max(driver, torch.mps.driver_allocated_memory())
record = {"epoch": epoch+1, "actual_stage_updates": updates, "train_objective": float(np.mean(losses)),
"calibration_passing_set_nll_by_temperature": {str(t): v for v, _, t in candidates}}
history.append(record)
# Each completed epoch survives cancellation/failure; never retry a prefix.
save(output / ("qat" if qat else "fp32") / f"epoch-{epoch+1:03d}.json", record, root)
if (epoch+1) % 20 == 0:
print(json.dumps({"arm": arm, "stage": "qat" if qat else "fp32", "epoch": epoch+1,
"actual_stage_updates": updates}), flush=True)
torch.mps.synchronize()
wall, after = time.perf_counter()-started, resource.getrusage(resource.RUSAGE_SELF)
cpu = after.ru_utime+after.ru_stime-before.ru_utime-before.ru_stime
core.require(updates == 800 and best is not None, "fixed actual per-stage update budget")
model.cpu().load_state_dict(best)
return model, {"optimizer_steps": updates, "epochs": 100, "training_function_groups": 1024,
"training_weighted_state_rows": len(train[0]), "initial_state_sha256": input_sha,
"selected_checkpoint_sha256": state_sha(best), "selected_epoch": selected[1],
"selected_temperature": selected[2], "selected_calibration_passing_set_nll": selected[0],
"selection": "minimum (calibration passing-set NLL, epoch, temperature) across all 100 epochs and four frozen temperatures",
"loop_wall_seconds": wall, "process_cpu_seconds_during_loop": cpu,
"process_cpu_percent_one_core_normalization": 100*cpu/wall,
"process_lifetime_peak_rss_bytes": after.ru_maxrss,
"mps_allocated_peak_sampled_bytes": allocated, "mps_driver_peak_sampled_bytes": driver,
"history": history}
def export(model, variant, directory, calibration, parity_inputs, x, selected_temperature):
core.export_model(model, variant, directory, 1., 1.)
cal_logits = core.exported_logits(directory, calibration[0])
if variant == "ptq_ternary":
temperature = min(TEMPERATURES, key=lambda t: (float((passing_nll(cal_logits, calibration[1], t)*calibration[2]).sum())/256, t))
else:
temperature = selected_temperature
metadata = read_json(directory / "model.json")
metadata.update(schema="gooo/tiny-three-choice-path-model/v1",
feature_version="triple_semantic_context_v3_joint_v1", temperature=temperature)
metadata.pop("confidence_threshold")
core.save_json(directory / "model.json", metadata)
vectors = x[[r["feature_row_index"] for r in parity_inputs]]
logits = core.exported_logits(directory, vectors)
probabilities = core.probabilities(logits, temperature)
parity = [{"variant": variant, "state_id": r["state_id"], "phase": r["phase"], "text": r["text"],
"features": vectors[i].tolist(), "logits": logits[i].tolist(), "probabilities": probabilities[i].tolist(),
"selected_mask": int(probabilities[i].argmax())} for i, r in enumerate(parity_inputs)]
pin = {"metadata_sha256": core.sha((directory / "model.json").read_bytes()),
"weights_sha256": core.sha((directory / "weights.bin").read_bytes()),
"packed_weights_bytes": (directory / "weights.bin").stat().st_size,
"temperature": temperature, "calibration_export_passing_set_nll": float((passing_nll(cal_logits, calibration[1], temperature)*calibration[2]).sum())/256}
core.require(pin["packed_weights_bytes"] == (74624 if variant == "fp32" else 3854), "exact separate three-choice weight ABI")
return pin, parity
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--prepared", type=Path, required=True)
parser.add_argument("--preparation-audit", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--source-revision", required=True)
args = parser.parse_args()
root = Path(__file__).resolve().parents[1]
head = subprocess.check_output(["git", "rev-parse", "HEAD"], cwd=root, text=True).strip()
core.require(head == args.source_revision and not subprocess.check_output(["git", "status", "--porcelain"], cwd=root).strip(), "clean exact offline optimizer source")
core.require(torch.backends.mps.is_available() and not args.output.exists(), "actual local MPS and fresh output")
core.require(args.output.resolve().parent == root / "runs" and args.output.name.startswith("own-three-"), "capped own-three raw phase")
core.require(core.sha((root / PROTOCOL).read_bytes()) == PROTOCOL_SHA, "immutable three-choice protocol")
core.FEATURES, core.HIDDEN, core.LIMIT, core.LABELS = 768, 24, 1600, [f"mask_{i}" for i in range(8)]
manifest, rows, x, parity_inputs = load_inputs(args.prepared, args.preparation_audit)
initial_state = state_from_blob(args.prepared / "initial-fp32.bin")
initial_sha = state_sha(initial_state)
core.require(initial_sha == manifest["files"]["initial-fp32.bin"]["sha256"], "Go common initialization preserved exactly")
pre = {"schema": "gooo/own-three-choice-training-preexecution/v1", "source_revision": head,
"sources_sha256": {name: core.sha((root / "training" / name).read_bytes())
for name in ("train_own_three_feedback_v1.py", "train_pilot_v2.py")},
"protocol_sha256": PROTOCOL_SHA, "student_states_sha256": STATES_SHA, "teacher_audit_sha256": AUDIT_SHA,
"prepared_manifest_sha256": core.sha((args.prepared / "manifest.json").read_bytes()),
"preparation_audit_sha256": core.sha(args.preparation_audit.read_bytes()),
"initial_state_sha256": initial_sha, "initial_seed": INITIAL_SEED, "shuffle_seed": SHUFFLE_SEED,
"arms": ARMS, "device": "mps", "torch": torch.__version__, "numpy": np.__version__,
"planned_optimizer_steps": 4800, "optimizer": "AdamW", "learning_rate": .001, "weight_decay": .01,
"temperatures": TEMPERATURES, "prior_retained_raw_bytes": retained_bytes(root), "whole_study_raw_cap_bytes": RAW_CAP,
"initialization": "Byte-exact new common Go initializer. Teacher observations only; no prior/pretrained/Laya student weights. QAT starts its own jointly calibration-selected FP checkpoint.",
"gpu_utilization_measured": False, "host_cpu_causal_delta_measured": False, "default_model_promoted": False}
save(args.output / "preexecution.json", pre, root)
reports, exports = {}, {}
total_updates = 0
try:
for arm in ARMS:
data = {s: arrays(rows, x, s, arm) for s in ("train", "calibration")}
fp, fp_report = fit(copy.deepcopy(initial_state), data, arm, False, args.output / arm, root)
total_updates += fp_report["optimizer_steps"]
qat, qat_report = fit(copy.deepcopy(fp.state_dict()), data, arm, True, args.output / arm, root)
total_updates += qat_report["optimizer_steps"]
reports[arm], exports[arm], parity = {"fp32": fp_report, "qat_ternary": qat_report}, {}, []
for variant, model, temperature in (("fp32", fp, fp_report["selected_temperature"]),
("ptq_ternary", fp, None),
("qat_ternary", qat, qat_report["selected_temperature"])):
core.require(retained_bytes(root) + 74624 + (64 << 10) <= RAW_CAP-(1 << 20),
"reserve complete model and failure metadata before export")
pin, records = export(model, variant, args.output / arm / "models" / variant,
data["calibration"], parity_inputs, x, temperature)
exports[arm][variant] = pin
parity.extend(records)
save_jsonl(args.output / arm / "go-parity.jsonl", parity, root)
save(args.output / arm / "training-report.json", reports[arm], root)
retained_bytes(root)
print(json.dumps({"completed_arm": arm, "actual_optimizer_updates": total_updates}), flush=True)
core.require(total_updates == 4800, "fixed total actual optimizer budget")
save(args.output / "report.json", {"schema": "gooo/own-three-choice-training-report/v1", "status": "TRAINED_AND_EXPORTED",
"optimizer_steps": total_updates, "preexecution_sha256": core.sha((args.output / "preexecution.json").read_bytes()),
"training": reports, "exports": exports, "default_model_promoted": False,
"scope": "All nine matched fresh three-choice FP/PTQ/QAT exports retained. Offline MPS optimization only; Go numerical parity, SDK continuation and native compilation remain separate uncompleted stages."}, root)
except BaseException as error:
save(args.output / "failure.json", {"status": "FAILED_PREFIX_RETAINED", "error_type": type(error).__name__,
"completed_stage_optimizer_updates": total_updates,
"scope": "Completed epoch receipts retain actual per-stage update counts. No restart, trimming, arm replacement or promotion."}, root)
raise
print(json.dumps({"status": "TRAINED_AND_EXPORTED", "optimizer_steps": total_updates, "model_exports": 9}), flush=True)
if __name__ == "__main__":
main()
|