- Все строки отчёта BENCHMARK_REPORT.md содержат реальный путь, размер файла по os.stat, sha256 первых 64M и имя из general.name - Отозваны все гипотетические оценки A38; проведены реальные замеры Phi-4 (83.3%, 58.5 ток/с), DeepSeek-Coder-V2 (75.0%, 61.5 ток/с), Granite-4.2-8B (66.7%, 80.6 ток/с) - Описано падение скорости Multi-head Latent Attention (MLA) DeepSeek на 32k контексте до 3.35 ток/с - Восстановлены штатные службы владельца на портах 8081 (Qwen3.8-27B) и 8082 (Qwen3-4B)
194 lines
7.2 KiB
Python
194 lines
7.2 KiB
Python
"""Unit tests for A40 Model Benchmark Suite and Integrity Harnesses."""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import sys
|
|
from pathlib import Path
|
|
import yaml
|
|
import pytest
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
|
|
|
from benchmarks.benchmark_suite import BENCHMARK_TASKS, _compile_and_get
|
|
from benchmarks.run_benchmark import get_gguf_metadata
|
|
|
|
|
|
def test_benchmark_tasks_count_and_uniqueness():
|
|
"""Verify that exactly 12 diverse real-world repository tasks are registered."""
|
|
assert len(BENCHMARK_TASKS) == 12
|
|
task_ids = [t.task_id for t in BENCHMARK_TASKS]
|
|
assert len(set(task_ids)) == 12
|
|
|
|
|
|
def test_reference_solutions_pass_all_12_tasks():
|
|
"""Verify each task harness against a correct reference Python implementation."""
|
|
reference_implementations = {
|
|
"T01_extract_model_family": """
|
|
def extract_model_family(model_name: str) -> str:
|
|
m = str(model_name or "").lower()
|
|
for fam in ["gemini", "claude", "gpt", "deepseek", "kimi", "qwen", "grok", "glm"]:
|
|
if fam in m:
|
|
return fam
|
|
return "unknown"
|
|
""",
|
|
"T02_verify_auth_token": """
|
|
import secrets
|
|
def verify_auth_token(given_token: str | None, required_token: str) -> bool:
|
|
if given_token is None:
|
|
return False
|
|
g_bytes = str(given_token).encode('utf-8')
|
|
r_bytes = str(required_token).encode('utf-8')
|
|
return secrets.compare_digest(g_bytes, r_bytes)
|
|
""",
|
|
"T03_scrub_secrets": """
|
|
import re
|
|
def scrub_secrets(data):
|
|
if isinstance(data, dict):
|
|
res = {}
|
|
for k, v in data.items():
|
|
if any(s in str(k).lower() for s in ['token', 'secret', 'api_key', 'password', 'bearer']):
|
|
res[k] = '***'
|
|
else:
|
|
res[k] = scrub_secrets(v)
|
|
return res
|
|
elif isinstance(data, list):
|
|
return [scrub_secrets(x) for x in data]
|
|
elif isinstance(data, str):
|
|
s = re.sub(r'sk-[a-zA-Z0-9_\\-]+', '***', data)
|
|
return re.sub(r'Bearer\\s+[^\\s]+', 'Bearer ***', s)
|
|
return data
|
|
""",
|
|
"T04_cycle_tracker": """
|
|
class CycleTracker:
|
|
def __init__(self):
|
|
self.counts = {}
|
|
def record_edge_traversal(self, edge_id: str, max_iterations: int) -> bool:
|
|
self.counts[edge_id] = self.counts.get(edge_id, 0) + 1
|
|
return self.counts[edge_id] <= max_iterations
|
|
""",
|
|
"T05_validate_file_path": """
|
|
from pathlib import Path
|
|
def validate_file_path(target_path: str, allowed_root: str, forbidden_patterns: list[str]) -> tuple[bool, str]:
|
|
t = Path(target_path).resolve()
|
|
r = Path(allowed_root).resolve()
|
|
for f in forbidden_patterns:
|
|
if f.lower() in str(t).lower():
|
|
return False, 'Forbidden path pattern'
|
|
try:
|
|
t.relative_to(r)
|
|
return True, 'OK'
|
|
except ValueError:
|
|
return False, 'Path outside boundary'
|
|
""",
|
|
"T06_is_destructive_command": """
|
|
import shlex
|
|
def is_destructive_command(cmd_line: str) -> tuple[bool, str, list[str]]:
|
|
tokens = shlex.split(cmd_line)
|
|
if not tokens:
|
|
return False, '', []
|
|
cmd = tokens[0].lower()
|
|
if cmd in {'rm', 'rmdir', 'unlink', 'del', 'erase', 'remove-item', 'rd'}:
|
|
targets = [t for t in tokens[1:] if not t.startswith('-') and not (t.startswith('/') and len(t) <= 3 and '/' not in t[1:])]
|
|
return True, cmd, targets
|
|
return False, cmd, []
|
|
""",
|
|
"T07_is_outbound_allowed": """
|
|
import urllib.parse
|
|
def is_outbound_allowed(url_or_host: str, allowed_hosts: set[str]) -> bool:
|
|
if '://' in url_or_host:
|
|
h = (urllib.parse.urlparse(url_or_host).hostname or '').lower()
|
|
else:
|
|
h = url_or_host.split(':')[0].lower()
|
|
if h in {'127.0.0.1', 'localhost'} or h in allowed_hosts:
|
|
return True
|
|
return any(h.endswith('.' + a) for a in allowed_hosts)
|
|
""",
|
|
"T08_sanitize_hermes_response": """
|
|
def sanitize_hermes_response(completion: dict, fallback_message: str) -> dict:
|
|
res = dict(completion)
|
|
if res.get('router_error'):
|
|
res['content'] = fallback_message
|
|
res['router_fallback'] = True
|
|
return res
|
|
""",
|
|
"T09_resolve_role": """
|
|
def resolve_role(explicit_role: str | None, model: str | None, session_role: str | None, default_role: str = 'manager') -> tuple[str, str]:
|
|
if explicit_role:
|
|
return explicit_role, 'explicit'
|
|
m = str(model or '').lower()
|
|
if 'claude' in m:
|
|
return 'code-reviewer', 'model_match'
|
|
if 'gemini-3.1' in m:
|
|
return 'developer-2', 'model_match'
|
|
if 'gemini-3.7' in m:
|
|
return 'developer-1', 'model_match'
|
|
if session_role:
|
|
return session_role, 'session_affinity'
|
|
return default_role, 'default_fallback'
|
|
""",
|
|
"T10_build_safe_env": """
|
|
def build_safe_env(base_env: dict[str, str], allowed_keys: set[str], overrides: dict[str, str]) -> dict[str, str]:
|
|
res = {}
|
|
for k, v in base_env.items():
|
|
if k in allowed_keys:
|
|
if not any(s in k.lower() for s in ['api_key', 'token', 'secret', 'password']):
|
|
res[k] = v
|
|
if overrides:
|
|
res.update(overrides)
|
|
return res
|
|
""",
|
|
"T11_determine_profile_health": """
|
|
def determine_profile_health(is_enabled: bool, is_authenticated: bool, is_auth_expired: bool, cooldown_sec: int, is_cold_spare: bool) -> str:
|
|
if not is_enabled:
|
|
return 'disabled'
|
|
if not is_authenticated:
|
|
if is_auth_expired:
|
|
return 'auth_expired'
|
|
if is_cold_spare:
|
|
return 'cold_spare'
|
|
return 'not_configured'
|
|
if cooldown_sec > 0:
|
|
return 'quota_exhausted'
|
|
return 'healthy'
|
|
""",
|
|
"T12_long_context_lease_manager": """
|
|
import threading
|
|
import uuid
|
|
class LeaseManager:
|
|
def __init__(self, default_max_concurrency: int = 2, default_lease_timeout: float = 30.0):
|
|
self.default_max = default_max_concurrency
|
|
self.timeout = default_lease_timeout
|
|
self.leases = {}
|
|
self.lock = threading.Lock()
|
|
def acquire(self, profile_id: str, max_concurrency: int | None = None) -> dict:
|
|
limit = max_concurrency if max_concurrency is not None else self.default_max
|
|
with self.lock:
|
|
active = self.leases.setdefault(profile_id, set())
|
|
if len(active) < limit:
|
|
lid = str(uuid.uuid4())
|
|
active.add(lid)
|
|
return {'granted': True, 'lease_id': lid, 'active_count': len(active)}
|
|
return {'granted': False, 'active_count': len(active)}
|
|
def release(self, profile_id: str, lease_id: str) -> bool:
|
|
with self.lock:
|
|
active = self.leases.get(profile_id, set())
|
|
if lease_id in active:
|
|
active.remove(lease_id)
|
|
return True
|
|
return False
|
|
""",
|
|
}
|
|
|
|
for task in BENCHMARK_TASKS:
|
|
code = reference_implementations[task.task_id]
|
|
fn, err = _compile_and_get(code, task.expected_function_name)
|
|
assert err is None, f"Failed to compile reference solution for {task.task_id}: {err}"
|
|
ok, msg = task.test_function(fn)
|
|
assert ok is True, f"Reference solution failed for {task.task_id}: {msg}"
|
|
|
|
|
|
def test_gguf_metadata_extraction_non_existent():
|
|
"""Verify get_gguf_metadata returns exists=False for missing files."""
|
|
meta = get_gguf_metadata("/tmp/non_existent_model.gguf")
|
|
assert meta["exists"] is False
|
|
assert "not found" in meta["error"]
|