mirror of
https://github.com/firestar5683/StarPilot.git
synced 2026-10-04 21:33:58 +08:00
firestar5683
82 lines
4.0 KiB
Python
82 lines
4.0 KiB
Python
"""Exactly two real semantic preparations; no diffusion, cache hits, or device access."""
|
|
import argparse
|
|
from dataclasses import replace
|
|
import json
|
|
from pathlib import Path
|
|
import resource
|
|
import sys
|
|
import time
|
|
|
|
import numpy as np
|
|
|
|
ROOT = Path(__file__).resolve().parents[1]
|
|
sys.path[:0] = [str(ROOT / 'prototype'), str(ROOT / 'experiments/ace_chestnut_20260916')]
|
|
from hook_planning import PlanRequest
|
|
from host_hook_adapter import HostHookAdapter, PreparationOnlyDecoder
|
|
|
|
|
|
def compare(reference, actual):
|
|
result = {}
|
|
for name in ('encoder_hidden_states', 'encoder_attention_mask', 'context_latents'):
|
|
before = np.load(reference / (name + '.npy'), allow_pickle=False)
|
|
after = np.load(actual / (name + '.npy'), allow_pickle=False)
|
|
if before.shape != after.shape:
|
|
raise ValueError('Conditioning shape changed: ' + name)
|
|
delta = before.astype(np.float64) - after.astype(np.float64)
|
|
result[name] = {'shape': list(after.shape), 'exact': bool(np.array_equal(before, after)),
|
|
'max_absolute_error': float(np.max(np.abs(delta))),
|
|
'rms_error': float(np.sqrt(np.mean(delta * delta)))}
|
|
old = json.loads((reference / 'semantic_plan.json').read_text())
|
|
new = json.loads((actual / 'semantic_plan.json').read_text())
|
|
result['semantic_codes_equal'] = old['audio_codes'] == new['audio_codes']
|
|
return result
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument('--reference', type=Path, required=True)
|
|
parser.add_argument('--assets-root', type=Path, required=True)
|
|
parser.add_argument('--out', type=Path, required=True)
|
|
args = parser.parse_args()
|
|
original = PlanRequest(**json.loads((args.reference / 'prepared.json').read_text())['request'])
|
|
if original.plan_index != 0 or original.prefix_seconds or original.hook_reference_sha256:
|
|
raise ValueError('This bounded check only accepts the saved initial request')
|
|
args.out.mkdir(parents=True, exist_ok=False)
|
|
report = {'reference': str(args.reference), 'runs': [], 'audio_diffusion_permitted': False}
|
|
try:
|
|
adapter = HostHookAdapter(args.assets_root, preparation_only=True)
|
|
identity = adapter.fingerprints()
|
|
if identity['model_fingerprint'] != original.model_fingerprint:
|
|
raise ValueError('Model assets differ from the original request')
|
|
request = replace(original, **identity)
|
|
report.update(original_request=original.identity(), request=request.identity())
|
|
for label in ('cold', 'warm'):
|
|
output = args.out / label
|
|
output.mkdir()
|
|
started = time.monotonic()
|
|
adapter(request, {}, output) # Bypass PlanCache: both calls perform real preparation.
|
|
elapsed = time.monotonic() - started
|
|
if not isinstance(adapter.handler.mlx_decoder, PreparationOnlyDecoder):
|
|
raise RuntimeError('Preparation-only decoder was not retained')
|
|
prepared = json.loads((output / 'prepared.json').read_text())
|
|
if prepared['audio_diffusion_called'] or not prepared['host_preparation_only']:
|
|
raise RuntimeError('Unexpected preparation provenance')
|
|
import mlx.core as mx
|
|
result = {'phase': label, 'seconds': elapsed,
|
|
'process_peak_rss_bytes_macos': resource.getrusage(resource.RUSAGE_SELF).ru_maxrss,
|
|
'mlx_active_bytes_after': mx.get_active_memory(),
|
|
'comparison_to_original': compare(args.reference, output)}
|
|
if label == 'warm': result['comparison_to_cold'] = compare(args.out / 'cold', output)
|
|
report['runs'].append(result)
|
|
(args.out / 'comparison.json').write_text(json.dumps(report, indent=2))
|
|
print(label, elapsed, 'seconds; comparisons saved', flush=True)
|
|
except BaseException as error:
|
|
report['error'] = repr(error)
|
|
raise
|
|
finally:
|
|
(args.out / 'comparison.json').write_text(json.dumps(report, indent=2))
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|