Digital Drosophila Archive
217 lines
12 KiB
Python
217 lines
12 KiB
Python
"""Bounded replay benchmark. No brain simulation, training, or weight installation."""
|
|
from __future__ import annotations
|
|
import argparse
|
|
from datetime import datetime, timezone
|
|
import hashlib
|
|
import json
|
|
import math
|
|
import os
|
|
from pathlib import Path
|
|
import statistics
|
|
import subprocess
|
|
import sys
|
|
import time
|
|
import traceback
|
|
import urllib.request
|
|
|
|
import numpy as np
|
|
from gpu_backend import CudaEligibility, prepare_events
|
|
|
|
REG_SHA = 'b63e6281b6505ca3b60acb467dd0b3fae274cf513dcbfd653649ba52e9ba9740'
|
|
IDENTITY_KEYS = ('kc', 'candidate_edge_id', 'candidate_sources', 'candidate_targets',
|
|
'candidate_compartments', 'base_candidate')
|
|
OUTPUTS = ('sampled_trace', 'final_trace', 'snapshots', 'delta_after')
|
|
|
|
def sha(path):
|
|
h = hashlib.sha256()
|
|
with Path(path).open('rb') as stream:
|
|
for chunk in iter(lambda: stream.read(8 << 20), b''):
|
|
h.update(chunk)
|
|
return h.hexdigest()
|
|
|
|
def save(path, value):
|
|
with Path(path).open('x', encoding='utf-8') as stream:
|
|
json.dump(value, stream, indent=2, sort_keys=True, allow_nan=False)
|
|
stream.write('\n')
|
|
|
|
def read(path):
|
|
return json.loads(Path(path).read_text())
|
|
|
|
def evidence(a):
|
|
return dict(dtype=a.dtype.str, shape=list(a.shape), sha256=hashlib.sha256(a.tobytes()).hexdigest())
|
|
|
|
def contained(root, relative):
|
|
p = root / relative
|
|
assert not p.is_symlink() and p.resolve().is_relative_to(root.resolve()), 'Escaping fixture path'
|
|
return p
|
|
|
|
def idle_queues():
|
|
result = {}
|
|
for port in (8188, 8189, 8190):
|
|
with urllib.request.urlopen(f'http://192.168.10.51:{port}/queue', timeout=5) as response:
|
|
q = json.load(response)
|
|
assert set(('queue_running', 'queue_pending')).issubset(q)
|
|
result[str(port)] = {key: len(q[key]) for key in ('queue_running', 'queue_pending')}
|
|
assert not any(result[str(port)].values()), 'ComfyUI has work; benchmark will not compete'
|
|
return result
|
|
|
|
def gpu_snapshot(uuid):
|
|
text = subprocess.check_output(['nvidia-smi', '--id=' + uuid,
|
|
'--query-gpu=uuid,name,pci.bus_id,memory.total,memory.used,utilization.gpu,driver_version',
|
|
'--format=csv,noheader,nounits'], text=True).strip()
|
|
fields = [s.strip() for s in text.split(',')]
|
|
assert fields[0] == uuid
|
|
return dict(zip(('uuid', 'name', 'pci_bus_id', 'memory_total_mib', 'memory_used_mib',
|
|
'utilization_percent', 'driver_version'), fields))
|
|
|
|
def stats(values):
|
|
return dict(n=len(values), min=min(values), median=statistics.median(values),
|
|
mean=statistics.mean(values), max=max(values))
|
|
|
|
def compare(result, arrays):
|
|
report = {}
|
|
for key in OUTPUTS:
|
|
actual, expected = getattr(result, key), arrays['reference_' + key]
|
|
exact = actual.dtype == expected.dtype and actual.shape == expected.shape and actual.tobytes() == expected.tobytes()
|
|
diff = np.abs(actual.astype(np.float64) - expected.astype(np.float64))
|
|
report[key] = dict(bitwise_exact=exact, actual=evidence(actual), expected=evidence(expected),
|
|
max_abs_error=float(np.max(diff)) if diff.size else 0.,
|
|
mean_abs_error=float(np.mean(diff)) if diff.size else 0.)
|
|
return report
|
|
|
|
def load_fixtures(manifests):
|
|
fixtures = []
|
|
common_identity = None
|
|
for manifest_path in manifests:
|
|
m = read(manifest_path)
|
|
assert m.get('complete') is True and m['registration_sha256'] == REG_SHA
|
|
root = manifest_path.parent
|
|
for f in m['fixtures']:
|
|
npz = contained(root, f['npz'])
|
|
command_path = contained(root, f['command'])
|
|
assert sha(npz) == f['npz_sha256'] and sha(command_path) == f['command_sha256']
|
|
with np.load(npz, allow_pickle=False) as z:
|
|
a = {k: z[k].copy() for k in z.files}
|
|
identity = {k: evidence(a[k]) for k in IDENTITY_KEYS}
|
|
if common_identity is None:
|
|
common_identity = identity
|
|
assert common_identity == identity, 'Full KC/mask/base identity differs across fixtures'
|
|
assert len(a['kc']) == 4064 and len(a['candidate_edge_id']) == 14232
|
|
source_pins = read(npz.parent / 'reference-sources.json')
|
|
for info in source_pins.values():
|
|
assert sha(info['path']) == info['sha256'], 'Reference source changed'
|
|
# Extract constants from the authenticated original bridge, without import/exec.
|
|
import ast
|
|
bridge = source_pins['eligibility_bridge_source']
|
|
assert bridge['sha256'] == '174fb8d14619189e74b095e77a6db18b14c93046dc431c0b55710a5aa2dd2e61'
|
|
constants = {}
|
|
for node in ast.parse(Path(bridge['path']).read_text()).body:
|
|
if isinstance(node, ast.Assign):
|
|
for target in node.targets:
|
|
if isinstance(target, ast.Name) and target.id in ('DT_MS', 'TAU_MS', 'ETA_MV', 'FLOOR_FRACTION', 'POLARITY'):
|
|
constants[target.id] = ast.literal_eval(node.value)
|
|
assert constants['POLARITY'] == -1
|
|
rule = dict(eta_mv=constants['ETA_MV'], floor_fraction=constants['FLOOR_FRACTION'],
|
|
decay=math.exp(-constants['DT_MS'] / constants['TAU_MS']))
|
|
fixtures.append(dict(group=root.name, record=f, arrays=a, command=read(command_path),
|
|
rule=rule, input_path=str(npz), manifest_sha256=sha(manifest_path)))
|
|
assert fixtures
|
|
return fixtures, common_identity
|
|
|
|
def main(args):
|
|
assert os.environ.get('CUDA_VISIBLE_DEVICES') == args.uuid, 'Use one explicit physical UUID per process'
|
|
assert args.output.is_absolute() and not args.output.exists()
|
|
args.output.mkdir()
|
|
started = time.perf_counter()
|
|
try:
|
|
queues_before = idle_queues()
|
|
gpu_before = gpu_snapshot(args.uuid)
|
|
fixtures, common_identity = load_fixtures(args.manifest)
|
|
backend = CudaEligibility(device_index=0)
|
|
cp = backend.cp
|
|
props = cp.cuda.runtime.getDeviceProperties(0)
|
|
bus = cp.cuda.runtime.deviceGetPCIBusId(0)
|
|
if isinstance(bus, bytes): bus = bus.decode()
|
|
assert bus.lower()[-10:] == gpu_before['pci_bus_id'].lower()[-10:], 'Physical GPU binding differs'
|
|
device = dict(**gpu_before, runtime_pci_bus_id=bus, compute_capability=backend.device.compute_capability,
|
|
initialization_seconds=backend.initialization_seconds, cupy_version=cp.__version__,
|
|
numpy_version=np.__version__, python_version=sys.version,
|
|
runtime_version=cp.cuda.runtime.runtimeGetVersion(), driver_version_runtime=cp.cuda.runtime.driverGetVersion(),
|
|
runtime_name=props['name'].decode() if isinstance(props['name'], bytes) else props['name'])
|
|
source = Path(__file__).resolve().parent
|
|
commit = subprocess.check_output(['git', '-C', str(source), 'rev-parse', 'HEAD'], text=True).strip()
|
|
assert subprocess.check_output(['git', '-C', str(source), 'status', '--porcelain'], text=True) == '', 'Candidate source worktree dirty'
|
|
save(args.output / 'authority.json', dict(device=device, queues_before=queues_before,
|
|
candidate_commit=commit, common_full_mapping_identity=common_identity,
|
|
fixture_manifests={str(p): sha(p) for p in args.manifest}, repeats=20, warmups=3,
|
|
all_source_files={p.name: sha(p) for p in sorted(source.iterdir()) if p.is_file()},
|
|
scope='eligibility/delta replay only; no brain, training, or installation',
|
|
original_experiment_resumed=False, created_utc=datetime.now(timezone.utc).isoformat()))
|
|
reports = []
|
|
peak_reserved = 0
|
|
for i, fixture in enumerate(fixtures):
|
|
assert time.perf_counter() - started < 900, 'Benchmark wall budget exhausted'
|
|
idle_queues()
|
|
a, c, rule = fixture['arrays'], fixture['command'], fixture['rule']
|
|
arrays_before = {k: evidence(v) for k, v in a.items()}
|
|
f_id = fixture['group'] + '--' + fixture['record']['fixture_id']
|
|
def run_once(resident=False):
|
|
begin = time.perf_counter()
|
|
prepared = prepare_events(a['step'], a['neuron_idx'], a['kc'], c['steps'],
|
|
a['reference_command_steps'], n_neurons=c['n_neurons'])
|
|
result = backend.run(prepared, a['candidate_sources'], a['candidate_compartments'],
|
|
c['compartment'], a['delta_before'], a['base_candidate'],
|
|
d=c['d_command'], frozen=c['frozen'], **rule,
|
|
kernel_warmup=3 if resident else 0, kernel_repeats=20 if resident else 1)
|
|
comparison = compare(result, a)
|
|
elapsed = time.perf_counter() - begin
|
|
if not all(r['bitwise_exact'] for r in comparison.values()):
|
|
save(args.output / (f_id + '-mismatch.json'), comparison)
|
|
raise AssertionError('GPU bitwise mismatch: ' + f_id)
|
|
return result, comparison, elapsed
|
|
cold, comp, cold_wall = run_once()
|
|
for _ in range(3):
|
|
run_once()
|
|
samples = []
|
|
for _ in range(20):
|
|
result, comp, elapsed = run_once()
|
|
samples.append(dict(**result.timing, replay_wall_including_preparation_and_comparison_seconds=elapsed))
|
|
resident, _, resident_wall = run_once(resident=True)
|
|
peak_reserved = max(peak_reserved, cp.get_default_memory_pool().total_bytes())
|
|
assert arrays_before == {k: evidence(v) for k, v in a.items()}, 'Input arrays mutated'
|
|
report = dict(fixture_id=f_id, input_path=fixture['input_path'], fixture_manifest_sha256=fixture['manifest_sha256'],
|
|
command=c, rule=rule, first_call_timing=cold.timing, first_call_wall_seconds=cold_wall,
|
|
raw_warm_samples=samples, resident_timing=resident.timing,
|
|
summaries={key: stats([s[key] for s in samples]) for key in
|
|
('preparation_seconds', 'kernel_seconds', 'end_to_end_seconds',
|
|
'replay_wall_including_preparation_and_comparison_seconds')},
|
|
comparison=comp, input_unchanged=True, all_repeats_bitwise_exact=True,
|
|
reference_timing=fixture['record'].get('reference_timing'), audit=result.audit)
|
|
save(args.output / (f_id + '.json'), report)
|
|
reports.append(report)
|
|
print(json.dumps(dict(fixture=f_id, exact=True,
|
|
median_ms=1000 * report['summaries']['replay_wall_including_preparation_and_comparison_seconds']['median'])), flush=True)
|
|
# Recheck original exported files at the end, not just at read time.
|
|
load_fixtures(args.manifest)
|
|
result = dict(status='PASS-ELIGIBILITY-UPDATE-REPLAY-EQUIVALENCE', fixtures=len(reports),
|
|
device=device, candidate_commit=commit, all_fixture_repeats_bitwise_exact=True,
|
|
observed_cuda_pool_reserved_peak_bytes=peak_reserved, gpu_after=gpu_snapshot(args.uuid),
|
|
queues_after=idle_queues(), wall_seconds=time.perf_counter()-started,
|
|
original_experiment_resumed=False, brain_simulation_tested=False,
|
|
full_training_speedup_measured=False, learning_success_claimed=False,
|
|
coverage_limit='available prefix + fixed synthetic edge cases, not all registered conditions',
|
|
files={p.name: sha(p) for p in sorted(args.output.iterdir()) if p.is_file()})
|
|
save(args.output / 'summary.json', result)
|
|
print(json.dumps(result), flush=True)
|
|
except Exception:
|
|
save(args.output / 'failure.json', dict(exit_code=1, traceback=traceback.format_exc(),
|
|
wall_seconds=time.perf_counter()-started, original_experiment_resumed=False))
|
|
raise
|
|
|
|
if __name__ == '__main__':
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument('--manifest', type=Path, action='append', required=True)
|
|
parser.add_argument('--output', type=Path, required=True)
|
|
parser.add_argument('--uuid', required=True)
|
|
main(parser.parse_args())
|