Skip to content

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++
5.5 KB
Raw Blame
"""The upstream comparison must reject mismatches and incomplete evidence."""
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_disabled import NAMES, SETTINGS, STATE_KEYS, common_args, state_hashes, verify


def fixture():
    requests = [dict(name=n, ids=[1, 2], max_new=16, expected='4' if n == 'short' else None) for n in NAMES]
    digest = hashlib.sha256(json.dumps(requests, sort_keys=True).encode()).hexdigest()
    baseline = dict(name='upstream', args=['--kv', 'int8'], info={key: 1 for key in SETTINGS},
                    request_digest=digest, cache_events=[], records=[])
    for name in NAMES:
        baseline['records'].append(dict(name=name, ids=[4], text='4', finish='stop', prompt_tokens=2,
                                        reused=2 if name in ('A-live', 'A-checkpoint') else 0,
                                        state={key: 'abc' for key in STATE_KEYS}))
    candidate = copy.deepcopy(baseline)
    candidate['name'] = 'candidate-1'
    candidate['info']['conversation_cache_mib'] = 0
    candidate['info']['conversation_disk_mib'] = 0
    return dict(arms=[baseline, candidate], requests=requests, request_digest=digest)


class DisabledCacheGateTest(unittest.TestCase):
    def test_reference_fixture(self):
        verify(fixture())

    def test_upstream_hash_schema(self):
        line = 'strata serve: STATE_HASH L=2 gdn=ab ple=cd tail=ef pooled=12 kv=34 mtp=56 stale=78 ple_prev=1,2'
        self.assertEqual(set(state_hashes(line)[0]), set(STATE_KEYS))
        self.assertEqual(state_hashes(line), state_hashes(line + ' dead=99'))
        # the candidate's pooled= is the upstream extent; pooled_full= (with the spare row) is not compared
        self.assertEqual(state_hashes(line), state_hashes(line + ' dead=99 pooled_full=ff', candidate=True))
        self.assertNotEqual(state_hashes(line), state_hashes(line.replace('pooled=12', 'pooled=ff') +
                                                           ' pooled_full=12', candidate=True))
        with self.assertRaises(AssertionError):
            state_hashes(line, candidate=True)
        with self.assertRaises(AssertionError):
            state_hashes(line.replace('kv=34 ', ''))

    def test_rejects_bad_evidence(self):
        mutations = [
            lambda r: r['arms'].pop(),
            lambda r: r['requests'].pop(),
            lambda r: r['requests'][0]['ids'].append(3),
            lambda r: r['arms'][1]['args'].append('--different'),
            lambda r: r['arms'][1].update(request_digest='different'),
            lambda r: r['arms'][1]['records'].pop(),
            lambda r: r['arms'][1]['info'].update(expert_slots=2),
            lambda r: r['arms'][1]['info'].pop('kv'),
            lambda r: r['arms'][1]['info'].update(conversation_cache_mib=8192),
            lambda r: r['arms'][1]['info'].pop('conversation_cache_mib'),
            lambda r: r['arms'][1]['info'].update(conversation_disk_mib=16384),
            lambda r: r['arms'][1]['cache_events'].append('disk cache: spilled'),
            lambda r: r['arms'][1]['records'][0].update(ids=[]),
            lambda r: r['arms'][1]['records'][0].update(ids=[5]),
            lambda r: r['arms'][1]['records'][0].update(finish='disconnect'),
            lambda r: r['arms'][1]['records'][0].update(prompt_tokens=1),
            lambda r: r['arms'][1]['records'][0]['state'].pop('gdn'),
            lambda r: r['arms'][1]['records'][0]['state'].update(kv='bad'),
            lambda r: r['arms'][1]['records'][0].update(reused=1),
            lambda r: [a['records'][0].update(text='5') for a in r['arms']],
            lambda r: [a['records'][2].update(reused=0) for a in r['arms']],
            lambda r: [a['records'][5].update(reused=2) for a in r['arms']],
        ]
        for mutation in mutations:
            with self.subTest(mutation=mutations.index(mutation)):
                result = fixture()
                mutation(result)
                with self.assertRaises(AssertionError):
                    verify(result)

    def test_optimized_python_still_rejects(self):
        code = ('from tools.test_conversation_cache_disabled import fixture, verify; '
                'r=fixture(); r["arms"][1]["records"][0]["ids"]=[5]; verify(r)')
        p = subprocess.run([sys.executable, '-O', '-c', code], capture_output=True, text=True)
        self.assertNotEqual(p.returncode, 0)
        self.assertIn('output differs', p.stderr)

    def test_cache_options_omitted_from_shared_arguments(self):
        cfg = {'args': ['--pack', '/model', '--conversation-cache-mib', '8192',
                        '--conversation-cache-disk', '/cache', '--kv', 'int8']}
        args = common_args(cfg, 4)
        self.assertEqual(args[:4], ['--pack', '/model', '--kv', 'int8'])
        self.assertFalse(any(x.startswith('--conversation-cache-') for x in args))

    def test_dry_run_does_not_load_or_create_output(self):
        with tempfile.TemporaryDirectory() as tmp:
            out = Path(tmp) / 'absent'
            p = subprocess.run([sys.executable, 'tools/conversation_cache_disabled.py', '--config', '/absent',
                                '--upstream', '/absent', '--candidate', '/absent', '--output', str(out)],
                               capture_output=True, text=True)
            self.assertEqual(p.returncode, 0, p.stderr)
            self.assertFalse(out.exists())


if __name__ == '__main__':
    unittest.main()