File size: 11,099 Bytes
47709ed | 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 | #!/usr/bin/env python3
"""Locked WikiText distillation/evaluation for QH-024 and controls."""
from __future__ import annotations
import argparse
import hashlib
import json
import math
import random
import sys
import time
from datetime import datetime, timezone
from pathlib import Path
import numpy as np
import torch
from transformers import AutoModelForMultimodalLM
PROJECT = Path(__file__).resolve().parents[2]
LOCK = PROJECT / "artifacts/wikitext-qh024-distillation-lock.json"
TRAIN_TOKENS = PROJECT / "datasets/processed/wikitext2_qwen38/train-tokens.pt"
VALIDATION_TOKENS = PROJECT / "datasets/processed/wikitext2_qwen38/validation-tokens.pt"
OUTPUT_DIR = PROJECT / "artifacts/wikitext_qh024"
sys.path.insert(0, str(PROJECT / "code/src"))
from quantum_qwen38.hyqut_replacement import QH024Config, freeze_and_replace_qh024
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument("--candidate", choices=["base", "svd512", "qh024", "cc024"], required=True)
parser.add_argument("--timing-context", choices=["exclusive", "shared"], default="shared")
return parser.parse_args()
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(8 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
def seed_everything(seed: int) -> None:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def token_block(tokens: torch.Tensor, index: int, length: int, device: torch.device) -> torch.Tensor:
start = index * length
return tokens[start : start + length].to(device=device, dtype=torch.long).unsqueeze(0)
def distill(model, replacement, tokens: torch.Tensor, lock: dict) -> dict:
protocol = lock["training_protocol"]
optimizer = torch.optim.AdamW(
[parameter for parameter in replacement.parameters() if parameter.requires_grad],
lr=float(protocol["learning_rate"]),
weight_decay=float(protocol["weight_decay"]),
)
losses, gradient_norms, seconds = [], [], []
model.eval()
model.config.use_cache = False
order = list(int(value) for value in lock["train_indices"])
random.Random(int(lock["seed"])).shuffle(order)
for index in order:
input_ids = token_block(tokens, index, int(lock["sequence_length"]), model.device)
replacement.set_capture_teacher_only(True)
with torch.inference_mode():
model(input_ids=input_ids, use_cache=False)
hidden_states, teacher_target = replacement.take_capture()
replacement.set_capture_teacher_only(False)
optimizer.zero_grad(set_to_none=True)
started = time.perf_counter()
student = replacement.student_forward(hidden_states)
denominator = teacher_target.float().square().mean().clamp_min(1.0e-8)
loss = (student.float() - teacher_target.float()).square().mean() / denominator
loss.backward()
gradient_norm = torch.nn.utils.clip_grad_norm_(
[parameter for parameter in replacement.parameters() if parameter.requires_grad],
float(protocol["gradient_clip_norm"]),
)
optimizer.step()
torch.cuda.synchronize(replacement.gamma.device)
losses.append(float(loss.detach()))
gradient_norms.append(float(gradient_norm))
seconds.append(time.perf_counter() - started)
del input_ids, hidden_states, teacher_target, student, denominator, loss
return {
"steps": len(losses),
"objective": protocol["objective"],
"learning_rate": float(protocol["learning_rate"]),
"loss_first_16_mean": float(np.mean(losses[:16])),
"loss_last_16_mean": float(np.mean(losses[-16:])),
"loss_min": float(np.min(losses)),
"gradient_norm_mean": float(np.mean(gradient_norms)),
"mean_residual_step_seconds_excluding_teacher_forward": float(np.mean(seconds)),
"total_residual_step_seconds_excluding_teacher_forward": float(np.sum(seconds)),
"losses": losses,
"frozen_parameter_grad_tensor_count": sum(
parameter.grad is not None for parameter in model.parameters() if not parameter.requires_grad
),
}
@torch.inference_mode()
def evaluate(model, tokens: torch.Tensor, indices: list[int], length: int) -> dict:
model.eval()
rows, seconds = [], []
for index in indices:
input_ids = token_block(tokens, index, length, model.device)
started = time.perf_counter()
loss = model(input_ids=input_ids, labels=input_ids, use_cache=False).loss
value = float(loss)
elapsed = time.perf_counter() - started
rows.append({
"block_index": index,
"mean_token_nll": value,
"predicted_tokens": length - 1,
"seconds": elapsed,
})
seconds.append(elapsed)
nll = float(np.mean([row["mean_token_nll"] for row in rows]))
return {
"metrics": {
"sequence_count": len(rows),
"predicted_token_count": len(rows) * (length - 1),
"mean_token_nll": nll,
"token_perplexity": math.exp(min(50.0, nll)),
"mean_sequence_seconds": float(np.mean(seconds)),
"p95_sequence_seconds": float(np.quantile(seconds, 0.95)),
},
"rows": rows,
}
def main() -> None:
args = parse_args()
if not torch.cuda.is_available():
raise RuntimeError("QH-024 formal runner is GPU-only")
lock = json.loads(LOCK.read_text(encoding="utf-8"))
if lock["status"] != "locked_before_qh024_qh_cc_base_performance_results":
raise RuntimeError("invalid QH-024 lock")
if lock["test_rows_inspected_or_used_for_fitness"] != 0:
raise RuntimeError("test leakage disclosure is nonzero")
seed_everything(int(lock["seed"]))
train_tokens = torch.load(TRAIN_TOKENS, map_location="cpu", weights_only=True)
validation_tokens = torch.load(VALIDATION_TOKENS, map_location="cpu", weights_only=True)
model_dir = Path((PROJECT / "records/active_model_path.txt").read_text().strip())
model = AutoModelForMultimodalLM.from_pretrained(
model_dir,
local_files_only=True,
dtype=torch.bfloat16,
device_map="balanced",
low_cpu_mem_usage=True,
attn_implementation="sdpa",
)
base_parameters = sum(parameter.numel() for parameter in model.parameters())
replacement = None
if args.candidate == "base":
for parameter in model.parameters():
parameter.requires_grad_(False)
injection = {
"candidate": "frozen-dense-base",
"base_model_parameters": base_parameters,
"deployed_model_parameters": base_parameters,
"net_model_parameter_reduction": 0,
"trainable_parameter_count": 0,
}
training = {"steps": 0}
else:
kind = "classical" if args.candidate == "cc024" else "quantum"
injection = freeze_and_replace_qh024(model, QH024Config(), core_kind=kind)
replacement = model.model.language_model.layers[7].self_attn.v_proj
if args.candidate == "svd512":
replacement.set_residual_enabled(False)
training = {"steps": 0, "role": "untrained causal zero-residual control"}
else:
training = distill(model, replacement, train_tokens, lock)
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
checkpoint_path = OUTPUT_DIR / f"wikitext-{args.candidate}-s{lock['seed']}-replacement.pt"
if replacement is not None and args.candidate in {"qh024", "cc024"}:
torch.save(
{
name: parameter.detach().cpu()
for name, parameter in replacement.named_parameters()
if parameter.requires_grad
},
checkpoint_path,
)
if replacement is not None:
replacement.strip_teacher()
eval_indices = [int(value) for value in lock["eval_indices"]]
evaluation = evaluate(model, validation_tokens, eval_indices, int(lock["sequence_length"]))
evaluation_zero = None
simulator_calls_before_zero = None
simulator_calls_after_zero = None
if replacement is not None and args.candidate in {"qh024", "cc024"}:
simulator_calls_before_zero = replacement.core.circuit_calls
replacement.set_residual_enabled(False)
evaluation_zero = evaluate(model, validation_tokens, eval_indices, int(lock["sequence_length"]))
simulator_calls_after_zero = replacement.core.circuit_calls
replacement.set_residual_enabled(True)
if simulator_calls_after_zero != simulator_calls_before_zero:
raise AssertionError("zero branch unexpectedly executed the residual core")
run_id = f"wikitext-{args.candidate}-s{lock['seed']}-n{lock['train_count']}-e{lock['eval_count']}"
payload = {
"status": "ok",
"run_id": run_id,
"candidate": args.candidate,
"model_dir": str(model_dir),
"timing_context": args.timing_context,
"dataset_lock": str(LOCK),
"dataset_lock_sha256": sha256(LOCK),
"dataset": {
"name": lock["dataset"],
"source_revision": lock["source_revision"],
"license": lock["license"],
"sequence_length": lock["sequence_length"],
"train_indices": lock["train_indices"] if training["steps"] else [],
"eval_indices": eval_indices,
"train_tokens_sha256": sha256(TRAIN_TOKENS),
"validation_tokens_sha256": sha256(VALIDATION_TOKENS),
"test_rows_inspected_or_used_for_fitness": 0,
},
"injection": injection,
"training": training,
"evaluation_active": evaluation,
"evaluation_zero_residual": evaluation_zero,
"simulator_calls_before_zero_evaluation": simulator_calls_before_zero,
"simulator_calls_after_zero_evaluation": simulator_calls_after_zero,
"replacement_audit_after_teacher_strip": replacement.audit() if replacement else None,
"checkpoint": str(checkpoint_path) if checkpoint_path.exists() else None,
"gpu_peak_allocated_bytes": [
torch.cuda.max_memory_allocated(index) for index in range(torch.cuda.device_count())
],
"checked_at": datetime.now(timezone.utc).isoformat(),
}
output = OUTPUT_DIR / f"{run_id}.json"
output.write_text(json.dumps(payload, indent=2), encoding="utf-8")
print(json.dumps({
"status": "ok",
"run_id": run_id,
"net_model_parameter_reduction": injection["net_model_parameter_reduction"],
"trainable_parameters": injection["trainable_parameter_count"],
"training_summary": {key: value for key, value in training.items() if key != "losses"},
"evaluation_active": evaluation["metrics"],
"evaluation_zero_residual": evaluation_zero["metrics"] if evaluation_zero else None,
"record": str(output),
}, indent=2))
if __name__ == "__main__":
main()
|