Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
"""Negative controls for the incremental capture model gate."""
import copy
import hashlib
import json
from pathlib import Path
import subprocess
import sys
import tempfile
import unittest
sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too
from tools.conversation_cache_growth import NAMES, SECRET, SETTINGS, STATE_KEYS, verify
def fixture():
lengths = [8, 2, 12, 2, 16, 2, 20, 2, 12, 2, 15, 2]
requests = [dict(name=name, ids=[1]*n, max_new=16, expected=name.startswith('A-'))
for name, n in zip(NAMES, lengths)]
digest = hashlib.sha256(json.dumps(requests, sort_keys=True).encode()).hexdigest()
full = dict(name='full', exe_sha256='fixed', args=['--conversation-cache-mib','1'], request_digest=digest,
info={key: 1 for key in SETTINGS}, draft_verifications=4,
parks=[dict(bytes=100,snapshot_bytes=100,reused_kv_bytes=0) for _ in range(3)],
records=[dict(name=name, ids=[1], text=SECRET, finish='stop', reused=1,
prompt_tokens=n, state={k:'abc' for k in STATE_KEYS}) for name,n in zip(NAMES,lengths)])
full['info'].update(conversation_cache_mib=1,conversation_cache_slots=4)
incremental = copy.deepcopy(full)
incremental['name'] = 'incremental'
for p in incremental['parks']:
p['reused_kv_bytes'] = 50
return dict(arms=[full,incremental],requests=requests,request_digest=digest,cache_mib=1)
class GrowthGate(unittest.TestCase):
def test_complete_evidence(self):
verify(fixture())
def test_incomplete_or_incorrect_evidence_fails(self):
mutations = [
lambda r:r['requests'].pop(),
lambda r:r['requests'][0]['ids'].append(2),
lambda r:r['arms'][1].update(exe_sha256='different'),
lambda r:r.update(cache_mib=2),
lambda r:r['arms'][1]['records'].pop(),
lambda r:r['arms'][1]['records'][2].update(ids=[2]),
lambda r:r['arms'][1]['records'][2].update(text='wrong'),
lambda r:r['arms'][1]['records'][2].update(reused=0),
lambda r:r['arms'][1]['records'][2].update(finish='cancel'),
lambda r:r['arms'][1]['records'][2]['state'].pop('kv'),
lambda r:r['arms'][1]['info'].update(expert_slots=2),
lambda r:r['arms'][1].update(draft_verifications=0),
lambda r:r['arms'][0]['parks'][0].update(reused_kv_bytes=1),
lambda r:[p.update(reused_kv_bytes=0) for p in r['arms'][1]['parks']],
lambda r:r['arms'][1]['parks'][0].update(bytes=2*1024*1024),
lambda r:r['arms'][1]['parks'][0].update(reused_kv_bytes=101),
]
for i, mutate in enumerate(mutations):
with self.subTest(case=i):
result = fixture()
mutate(result)
with self.assertRaises(AssertionError):
verify(result)
def test_optimized_python_rejects_missing_reuse(self):
code = ('from tools.test_conversation_cache_growth import fixture,verify; '
'r=fixture(); [p.update(reused_kv_bytes=0) for p in r["arms"][1]["parks"]]; verify(r)')
run = subprocess.run([sys.executable,'-O','-c',code],capture_output=True,text=True)
self.assertNotEqual(run.returncode,0)
self.assertIn('did not reuse KV bytes',run.stderr)
def test_dry_run_leaves_server_and_output_alone(self):
with tempfile.TemporaryDirectory() as tmp:
output = Path(tmp)/'absent'
run = subprocess.run([sys.executable,'tools/conversation_cache_growth.py','--config','/absent',
'--engine','/absent','--output',str(output)],capture_output=True,text=True)
self.assertEqual(run.returncode,0,run.stderr)
self.assertFalse(output.exists())
if __name__ == '__main__':
unittest.main()