Artvasil09
ce06659ab2
- Offline neural worker based on llama.cpp (fortune-engine.exe) with local Qwen3-4B model support - Native assembly Tarot 3-card spread divination and teacher grade divination mode - Text input dialog in MODE_FORTUNE communicating over anonymous pipes - Clean process lifecycle management, WM_CHAR dispatch and fallback handling - Test suites and packaging tools for offline release
78 lines
3.4 KiB
Python
78 lines
3.4 KiB
Python
"""Exercise the actual offline inference child through its bounded binary protocol."""
|
||
from pathlib import Path
|
||
import argparse
|
||
import json
|
||
import struct
|
||
import subprocess
|
||
import time
|
||
|
||
ROOT = Path(__file__).resolve().parents[2]
|
||
OUT = ROOT / 'build/fortune-ai'
|
||
|
||
|
||
def exact(stream, size):
|
||
data = bytearray()
|
||
while len(data) < size:
|
||
chunk = stream.read(size - len(data))
|
||
assert chunk, 'worker closed an incomplete response'
|
||
data.extend(chunk)
|
||
return bytes(data)
|
||
|
||
|
||
def response(worker):
|
||
magic, status, length = struct.unpack('<III', exact(worker.stdout, 12))
|
||
assert magic == 0x31524346 and length <= 8192
|
||
return status, exact(worker.stdout, length).decode('cp1251')
|
||
|
||
|
||
def main():
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument('--short', action='store_true')
|
||
args = parser.parse_args()
|
||
OUT.mkdir(exist_ok=True)
|
||
manifest = json.loads((ROOT / 'ai/model.json').read_text('utf8'))
|
||
assert (ROOT / 'ai' / manifest['file']).stat().st_size == manifest['bytes']
|
||
questions = [
|
||
('Стоит ли мне начать учиться рисованию?', [0, 23, 17]),
|
||
('А что мне поможет не бросить это через неделю?', [8, 36, 21]),
|
||
('Поставить реальному студенту 8 или 9 по его имени и году рождения?', [11, 4, 5]),
|
||
]
|
||
expected = [('шут', 'маг', 'звезд'), ('сил', 'умеренн', 'мир')]
|
||
worker = subprocess.Popen([str(ROOT / 'ai/fortune-engine.exe')], cwd=OUT,
|
||
stdin=subprocess.PIPE, stdout=subprocess.PIPE,
|
||
stderr=subprocess.DEVNULL, creationflags=0x08000000)
|
||
started = time.monotonic()
|
||
results = []
|
||
try:
|
||
status, text = response(worker)
|
||
assert (status, text) == (0, ''), (status, text)
|
||
ready = time.monotonic() - started
|
||
for index, (question, cards) in enumerate(questions[:1 if args.short else 3]):
|
||
before = time.monotonic()
|
||
payload = question.encode('cp1251')
|
||
worker.stdin.write(struct.pack('<IIIII', 0x31514346, len(payload), *cards) + payload)
|
||
worker.stdin.flush()
|
||
status, text = response(worker)
|
||
assert status in (1, 3), (status, text)
|
||
assert len(text) > 160, text
|
||
assert sum('а' <= c.lower() <= 'я' for c in text) > 80, text
|
||
assert '<think>' not in text and '<|im_' not in text
|
||
if index < len(expected):
|
||
assert all(card in text.lower().replace('ё', 'е') for card in expected[index]), text
|
||
results.append({'question': question, 'cards': cards, 'status': status,
|
||
'seconds': round(time.monotonic() - before, 2), 'answer': text})
|
||
# Reject invalid length before allocating or reading an unbounded body.
|
||
worker.stdin.write(struct.pack('<IIIII', 0x31514346, 0xffffffff, 0, 1, 2))
|
||
worker.stdin.flush()
|
||
assert worker.wait(timeout=15) == 2
|
||
finally:
|
||
if worker.poll() is None:
|
||
worker.kill()
|
||
worker.communicate(timeout=15)
|
||
(OUT / 'inference-results.json').write_text(json.dumps(
|
||
{'ready_seconds': round(ready, 2), 'results': results}, ensure_ascii=False, indent=2), encoding='utf8')
|
||
print(f'PASS: real offline Russian inference, supplied cards, bounded IPC; ready in {ready:.1f}s')
|
||
|
||
|
||
if __name__ == '__main__':
|
||
main()
|