Seal exact replay evidence and paired stage performance without training claims
This commit is contained in:
Digital Drosophila Archive
@@ -0,0 +1,194 @@
|
||||
"""Seal bounded engineering replay evidence; never resume scientific work."""
|
||||
from datetime import datetime, timezone
|
||||
import hashlib
|
||||
import json
|
||||
from pathlib import Path
|
||||
import statistics
|
||||
import subprocess
|
||||
|
||||
ROOT = Path('/home/qishizhe/phase4ar-pipeline-acceleration-20261001')
|
||||
SOURCE = Path(__file__).absolute().parent
|
||||
WORKTREE = SOURCE.parents[2]
|
||||
|
||||
|
||||
def require(value, message):
|
||||
if not value:
|
||||
raise ValueError(message)
|
||||
|
||||
|
||||
def read(path):
|
||||
return json.loads(path.read_text())
|
||||
|
||||
|
||||
def sha(path):
|
||||
h = hashlib.sha256()
|
||||
with path.open('rb') as stream:
|
||||
for block in iter(lambda: stream.read(8 << 20), b''):
|
||||
h.update(block)
|
||||
return h.hexdigest()
|
||||
|
||||
|
||||
def save(path, value):
|
||||
with path.open('x') as stream:
|
||||
json.dump(value, stream, indent=2, sort_keys=True, allow_nan=False)
|
||||
stream.write('\n')
|
||||
|
||||
|
||||
def stats(values):
|
||||
return dict(count=len(values), median=statistics.median(values), mean=statistics.mean(values),
|
||||
minimum=min(values), maximum=max(values), total=sum(values))
|
||||
|
||||
|
||||
def successful_job(label):
|
||||
record = read(ROOT / 'jobs' / (label + '.exit.json'))
|
||||
require(record['observed_exit'] is True and record['exit_code'] == 0
|
||||
and not record.get('stop_reason') and not record['timed_out'], 'Job did not exit successfully: ' + label)
|
||||
return record
|
||||
|
||||
|
||||
def check_inventory(root, inventory):
|
||||
for rel, item in inventory.items():
|
||||
path = root / rel
|
||||
require(path.resolve().is_relative_to(root) and not path.is_symlink(), 'Unsafe artifact path')
|
||||
expected = item if isinstance(item, str) else item['sha256']
|
||||
require(sha(path) == expected, 'Artifact changed: ' + str(path))
|
||||
if isinstance(item, dict) and 'bytes' in item:
|
||||
require(path.stat().st_size == item['bytes'], 'Artifact size changed')
|
||||
|
||||
|
||||
def main():
|
||||
require(ROOT.is_dir() and not ROOT.is_symlink(), 'New experiment root required')
|
||||
require(not (ROOT / 'final-summary.json').exists(), 'Finalization must not overwrite prior results')
|
||||
jobs = {name: successful_job(name) for name in
|
||||
('cpu-unit-tests', 'brain-unit-tests', 'cpu-profile-01', 'brain-replay-01')}
|
||||
before, after = [read(ROOT / ('integrity-' + stage + '.json')) for stage in ('before', 'after')]
|
||||
require(before['status'] == after['status'] == 'PASS', 'Pre/post integrity failed')
|
||||
require(after['original']['status'] == 'PASS' and after['original']['files_checked'] == 64034
|
||||
and not after['original']['file_mismatches'] and not after['original']['source_mismatches'], 'Science integrity failed')
|
||||
cpu = read(ROOT / 'cpu-profile/report.json')
|
||||
brain = read(ROOT / 'brain-replay/summary.json')
|
||||
require(cpu['complete'] and cpu['fixtures'] == 14 and cpu['all_child_exits_observed_zero']
|
||||
and cpu['all_full_array_reference_comparisons_passed'] and cpu['all_original_receipts_exact']
|
||||
and cpu['full_AQ_validates_actual_second_pass_E_outputs'], 'CPU regression incomplete')
|
||||
require(brain['complete'] and brain['all_bitwise_exact'] and brain['fresh_trial_count'] == 14
|
||||
and brain['resident_trial_count'] == 28 and brain['actual_child_processes'] == 15, 'Brain regression incomplete')
|
||||
check_inventory(ROOT / 'cpu-profile', cpu['files'])
|
||||
check_inventory(ROOT / 'brain-replay', brain['files'])
|
||||
fresh = [read(ROOT / ('brain-replay/sessions/fresh-%02d.json' % i))['records'][0] for i in range(14)]
|
||||
resident = read(ROOT / 'brain-replay/sessions/resident.json')['records']
|
||||
for record in fresh + resident:
|
||||
require(record['exact'] and record['input_pins_verified_before_and_after']
|
||||
and record['neural_state_fresh_each_call'] and not record['generated_delta_installed'], 'Invalid replay trial')
|
||||
paired = []
|
||||
for first in fresh:
|
||||
others = [r for r in resident if r['fixture_id'] == first['fixture_id']]
|
||||
require(len(others) == 2 and {r['direction'] for r in others} == {'forward', 'reverse'}, 'Incomplete fixed ordering')
|
||||
fresh_runtime = first['timings']['runtime_and_builtin_output_seconds']
|
||||
warm_runtime = statistics.median(r['timings']['runtime_and_builtin_output_seconds'] for r in others)
|
||||
warm_wall = statistics.median(r['trial_wall_seconds'] for r in others)
|
||||
paired.append(dict(fixture_id=first['fixture_id'], fixed_index=first['fixed_index'],
|
||||
first_fixture_cold_build_and_context=first['fixed_index'] == 0,
|
||||
fresh_runtime_seconds=fresh_runtime, resident_runtime_seconds=warm_runtime,
|
||||
runtime_ratio=fresh_runtime / warm_runtime,
|
||||
fresh_trial_wall_seconds=first['trial_wall_seconds'], resident_trial_wall_seconds=warm_wall,
|
||||
trial_wall_ratio=first['trial_wall_seconds'] / warm_wall,
|
||||
fresh_simulation_seconds=first['timings']['original_diagnostic_simulation_seconds'],
|
||||
resident_simulation_seconds=statistics.median(r['timings']['original_diagnostic_simulation_seconds'] for r in others)))
|
||||
# Exclusion is by fixed initialization position, never by observed outcome.
|
||||
# All 42 trials remain in correctness evidence and the full paired table.
|
||||
steady = paired[1:]
|
||||
brain_timing = {name: stats([p[name] for p in steady]) for name in (
|
||||
'fresh_runtime_seconds', 'resident_runtime_seconds', 'runtime_ratio',
|
||||
'fresh_trial_wall_seconds', 'resident_trial_wall_seconds', 'trial_wall_ratio',
|
||||
'fresh_simulation_seconds', 'resident_simulation_seconds')}
|
||||
cold_E = cpu['full_E']['cold_fixed_first']
|
||||
warm_E = [r for r in cpu['full_E']['resident']
|
||||
if r['round_index'] == 1 and r['fixture_id'] == cold_E['fixture_id']]
|
||||
require(len(warm_E) == 1, 'Missing same-fixture E timing')
|
||||
e_pair = dict(fixture_id=cold_E['fixture_id'],
|
||||
cold_API_seconds=cold_E['original_api_seconds'], warm_API_seconds=warm_E[0]['original_api_seconds'],
|
||||
API_ratio=cold_E['original_api_seconds'] / warm_E[0]['original_api_seconds'],
|
||||
sample_warning='One fixed-first cold sample; not a 14-fixture matched cold E study')
|
||||
commit = subprocess.check_output(['git', '-C', str(WORKTREE), 'rev-parse', 'HEAD'], text=True).strip()
|
||||
git_status = subprocess.check_output(['git', '-C', str(WORKTREE), 'status', '--porcelain'], text=True)
|
||||
require(not git_status, 'Candidate worktree must be committed before finalization')
|
||||
result = dict(status='PASS-PERSISTENT-REPLAY-EQUIVALENCE', created_utc=datetime.now(timezone.utc).isoformat(),
|
||||
source_commit=commit, registration_sha256=cpu['registration_sha256'],
|
||||
original_source_commit=cpu['original_source_commit'], fixture_manifest_sha256=cpu['fixture_manifest_sha256'],
|
||||
jobs=jobs, cpu_summary=cpu['summary'], full_E_same_fixture_API_comparison=e_pair,
|
||||
brain_matched_pairs=paired, brain_steady_13_fixture_summary=brain_timing,
|
||||
first_brain_fixture=paired[0],
|
||||
full_brain_regression=dict(fresh=14, resident_forward=14, resident_reverse=14,
|
||||
arrays_per_trial=10, all_bitwise_exact=True, actual_child_exits_observed_zero=15),
|
||||
cpu_regression=dict(cold_U=14, resident_update_plus_independent_verification=42,
|
||||
full_E_cold=1, full_E_resident=28, full_AQ_on_new_E_outputs=14, host_tests=17),
|
||||
integrity=after, scientific_AR_resumed=False, training_epochs_added=0, probes_added=0,
|
||||
production_runtime_promoted=False, scientific_learning_success_claimed=False,
|
||||
limitations=[
|
||||
'Fixed 14 real committed-prefix fixtures only; not the full scientific design.',
|
||||
'Fresh means fresh interpreter, not flushed filesystem caches.',
|
||||
'Brain uses unchanged original 4090 runtime; whole-brain V100 compatibility remains unestablished.',
|
||||
'Original GPU tensors/graph are still allocated and uploaded each trial; no numerical loop rewrite.',
|
||||
'Brain runtime timings include its original output I/O. Trial wall also includes extra audit checks.',
|
||||
'Resident trial audit additionally checks the current fresh reference, making trial-wall boundaries unequal.',
|
||||
'Paired steady timing excludes fixed fixture 0 cold initialization in both lanes; correctness excludes none.',
|
||||
'Independent measured stages are not a combined training pipeline; no end-to-end speedup or revised full-run ETA.',
|
||||
'Full independent AQ still uses a fresh subprocess and all original checks.',
|
||||
'The original stopped scientific run remains at 1033 CPU-verified epochs and 138 probes.',
|
||||
'Integrity covers saved Phase4AR files/sources and previous acceleration files; not a new Phase1-4AQ numerical regression.',
|
||||
'Expected timeout self-test intentionally exits nonzero; it is not a scientific-run failure.'
|
||||
])
|
||||
save(ROOT / 'final-summary.json', result)
|
||||
s = cpu['summary']
|
||||
b = brain_timing
|
||||
report = f'''# 常驻进程工程回放结果
|
||||
|
||||
状态：`PASS-PERSISTENT-REPLAY-EQUIVALENCE`。仅限这次固定回放，不代表学习成功或正式运行时升级。
|
||||
|
||||
## 实测
|
||||
|
||||
| 测量范围 | 新进程 | 常驻进程 | 说明 |
|
||||
|---|---:|---:|---|
|
||||
| 原始 CPU 更新 API | {s['cold_U_API_seconds']['median']:.4f} s | {s['resident_warm_U_API_seconds']['median']:.4f} s | 新进程首次 API 包含导入等开销 |
|
||||
| 完整更新 API，同一固定案例 | {e_pair['cold_API_seconds']:.4f} s | {e_pair['warm_API_seconds']:.4f} s | 仅一组冷启动对照 |
|
||||
| 全脑原始运行函数，13 组配对中位数 | {b['fresh_runtime_seconds']['median']:.4f} s | {b['resident_runtime_seconds']['median']:.4f} s | 含原有输出；排除固定首例的初始化 |
|
||||
|
||||
完整更新常驻 14 组中位耗时：{s['original_full_E_warm_roundtrip_seconds']['median']:.4f} s。
|
||||
完整独立验证仍需：{s['original_full_AQ_subprocess_wall_seconds']['median']:.4f} s（14 组中位数）。
|
||||
全脑运行函数同案例配对倍率中位数：{b['runtime_ratio']['median']:.3f}×，不是整轮训练加速倍率。
|
||||
|
||||
## 一致性与保护
|
||||
|
||||
- 17 项主机测试通过。
|
||||
- CPU：14 次冷启动更新、42 次独立双进程更新/重算、29 次完整更新、14 次完整独立验证通过。
|
||||
- 全脑：14 次新进程、正序 14 次与逆序 14 次常驻回放；所有脉冲、状态、延迟队列和请求等一致，10 个数组逐位相同。
|
||||
- 64,034 个原实验文件与 16 个注册源码重新校验通过；之前 659 个加速产物和 13 个源码保持不变。
|
||||
- 原 Phase4AR 仍停止于 1,033 轮 CPU 校验训练与 138 次 probe，没有新增训练或安装新权重。
|
||||
|
||||
## 边界与下一步
|
||||
|
||||
本次全脑测试使用原 RTX 4090 环境，没有证明全脑已能迁移到 V100。此前 V100 约 3.4 ms 的成绩仅属于 eligibility/update 模块。
|
||||
常驻进程复用的是解释器、模块和 CUDA 上下文；每轮神经状态仍由原始函数重新分配。
|
||||
当前仍保留完整验证、图和权重上传、记录写盘。下一步应在新的可审计流水线中验证整轮开销，不能把本报告的独立阶段相加当作实测训练速度。
|
||||
|
||||
源码提交：`{commit}`。完整逐例耗时、实际进程退出、来源校验见 `final-summary.json` 和各 lane 报告。
|
||||
'''
|
||||
with (ROOT / 'RESULTS.md').open('x') as stream:
|
||||
stream.write(report)
|
||||
files = {str(path.relative_to(ROOT)): dict(bytes=path.stat().st_size, sha256=sha(path))
|
||||
for path in sorted(ROOT.rglob('*')) if path.is_file()}
|
||||
source_files = {path.name: sha(path) for path in sorted(SOURCE.iterdir()) if path.is_file()}
|
||||
manifest = dict(schema='conditioning-persistent-replay-artifacts-v1', source_commit=commit,
|
||||
root=str(ROOT), files=files, source_files=source_files,
|
||||
self_hash_policy='This manifest excludes itself and SHA256SUMS; SHA256SUMS includes the manifest.')
|
||||
save(ROOT / 'artifact-manifest.json', manifest)
|
||||
with (ROOT / 'SHA256SUMS').open('x') as stream:
|
||||
for rel, item in files.items():
|
||||
stream.write(item['sha256'] + ' ' + rel + '\n')
|
||||
stream.write(sha(ROOT / 'artifact-manifest.json') + ' artifact-manifest.json\n')
|
||||
print(json.dumps(dict(status=result['status'], files=len(files), source_commit=commit,
|
||||
total_bytes=sum(v['bytes'] for v in files.values()), brain=brain_timing, CPU=cpu['summary']), indent=2))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
Reference in New Issue
Block a user