Function generate_tests

scripts/generate_config_database_tests.py:178–234

(test_list_path: Path = TEST_LIST_PATH, test_config_dir: Path = PERF_SANITY_DIR)

def generate_tests(test_list_path: Path = TEST_LIST_PATH, test_config_dir: Path = PERF_SANITY_DIR):
    test_list_path.parent.mkdir(parents=True, exist_ok=True)

    all_recipes = RecipeList.from_yaml(DATABASE_LIST_PATH)
    recipes = filter_to_key_recipes(all_recipes)
    print(f"Selected {len(recipes)} key recipes from {len(all_recipes)} total")

    gpu_groups = group_recipes_by_gpu(recipes)
    condition_entries = []
    config_files = {}

    for gpu_name in sorted(gpu_groups.keys()):
        gpu_recipes = gpu_groups[gpu_name]
        config_name = f"config_database_{gpu_name.lower()}"
        config_path = test_config_dir / f"{config_name}.yaml"

        aggr_config = generate_aggr_config(gpu_recipes)
        config_content = yaml.dump(
            aggr_config, default_flow_style=False, sort_keys=False, width=120
        )

        with open(config_path, "w", encoding="utf-8") as f:
            f.write(config_content)
        print(f"Generated {config_path}")

        config_files[config_path] = config_content

        # Generate condition entries grouped by num_gpus
        num_gpus_groups = group_recipes_by_num_gpus(gpu_recipes)
        for num_gpus in sorted(num_gpus_groups.keys()):
            server_names = [generate_server_name(r) for r in num_gpus_groups[num_gpus]]
            entry = generate_condition_entry(gpu_name, num_gpus, config_name, server_names)
            condition_entries.append(entry)

    test_list = {
        "version": "0.0.1",
        "llm_config_database": condition_entries,
    }

    header = """# ===============================================================================
# Config Database Performance Tests (AUTO-GENERATED)
# ===============================================================================
# Generated by: scripts/generate_config_database_tests.py
#
# These tests use scenario-only matching (match_mode: scenario) for baselines.
# Baselines are matched by (model, gpu, isl, osl, concurrency, num_gpus) instead
# of full config fields, allowing configs to evolve while maintaining comparison.
#
# To regenerate:
#   python scripts/generate_config_database_tests.py
# ===============================================================================

"""
    with open(test_list_path, "w", encoding="utf-8") as f:
        f.write(header)
        yaml.dump(test_list, f, default_flow_style=False, sort_keys=False, width=120)
    print(f"Generated {test_list_path}")