refactor(cli): одиночный файл и батч переведены на module выполнения

CLI строит один ExecutionRequest, создаёт Transcriber после prescan и получает результат с ExecutionInfo: разрешение auto, device-aware умолчания, проверка пословного контракта, строка устройства для шапки и fallback ушли за seam. --verbose печатает движок, потоки и runtime_info(). Удалены load_model, _transcribe_file, TranscribeFileResult, apply_device_defaults и тесты переноса полей; сценарные тесты CLI идут через CliRunner с fake adapter на get_backend. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-22 00:05:27 +03:00
1 parent e22b2f1bae
commit 0f8ecb5bd3
@@ -167,7 +167,6 @@ class FasterWhisperBackend:
        except Exception as exc:  # noqa: BLE001 — диагностика не должна ронять запуск
            cuda_devices = f"недоступно ({exc})"
        return {
-            "engine": "faster-whisper",
            "faster_whisper": package_version("faster-whisper"),
            "ctranslate2": package_version("ctranslate2"),
            "cuda_devices": cuda_devices,
@@ -229,7 +229,6 @@ class OnnxAsrBackend:

        configured = ", ".join(self._providers)
        return {
-            "engine": "onnx-asr",
            "onnxruntime": package_version("onnxruntime"),
            "onnx_asr": package_version("onnx-asr"),
            "available_providers": ", ".join(onnxruntime.get_available_providers()),
@@ -209,7 +209,6 @@ class OpenVINOBackend:
        from openvino import Core

        return {
-            "engine": "openvino",
            "openvino": package_version("openvino"),
            "available_devices": ", ".join(Core().available_devices),
            "device": self.actual_ov_device or "",
@@ -13,8 +13,6 @@ from rich.status import Status

 from .config import (
    CliValues,
-    ResolvedConfig,
-    apply_device_defaults,
    load_config,
    resolve_defaults,
 )
@@ -39,11 +37,12 @@ from .quality import (
 )
 from .speaker_diarizer import SpeakerDiarizer, load_speaker_diarizer
 from .transcriber import (
+    ExecutionInfo,
+    ExecutionRequest,
    Segment,
+    Transcriber,
    TranscribeResult,
-    _transcribe_file,
    cuda_error_hint,
-    load_model,
 )
 from .types import (
    UNKNOWN_LANGUAGE,
@@ -53,10 +52,7 @@ from .types import (
 )
 from .utils import (
    build_output_path,
-    detect_device,
    expand_globs,
-    get_gpu_name,
-    get_intel_gpu_name,
    has_existing_transcript,
    validate_input_file,
 )
@@ -95,21 +91,6 @@ def _with_cli_warning_renderer(
    return wrapped


-def _format_device_info(device_used: str) -> str:
-    """Формирует строку устройства для шапки транскрипта."""
-    if device_used == "cuda":
-        gpu_name = get_gpu_name()
-        return f"CUDA ({gpu_name or 'Unknown GPU'})"
-    if device_used == "openvino-gpu":
-        gpu_name = get_intel_gpu_name()
-        return f"OpenVINO ({gpu_name or 'Intel GPU'})"
-    if device_used in ("openvino", "openvino-cpu"):
-        return "OpenVINO (CPU)"
-    if device_used == "onnx":
-        return "ONNX (CPU)"
-    return "CPU"
-
-
 def _format_language_mode(requested_language: str, result: TranscribeResult) -> str:
    """Описывает источник языка, не выдавая профиль модели за детектор."""
    if requested_language != "auto":
@@ -348,7 +329,7 @@ def main(
        )
        raise SystemExit(2)

-    resolved_device: str | None = None
+    requested_device: str | None = None
    try:
        config = load_config()
        cli_values: CliValues = {
@@ -360,11 +341,19 @@ def main(
        }
        defaults = resolve_defaults(cli_values, config)
        diarize_enabled = defaults["diarize"] is True or speakers is not None
+        requested_device = defaults["device"]

-        resolved_device = detect_device(defaults["device"])
-        defaults = apply_device_defaults(defaults, resolved_device, cli_values, config)
-
-        ct_explicit = compute_type is not None or "compute_type" in config
+        # Явные model/compute_type; умолчания исполнения разрешает module.
+        request = ExecutionRequest(
+            device=requested_device,
+            model=model if model is not None else config.get("model"),
+            compute_type=(
+                compute_type if compute_type is not None else config.get("compute_type")
+            ),
+            language=defaults["language"],
+            cpu_threads=threads,
+            require_word_timestamps=diarize_enabled,
+        )

        expanded = expand_globs(files)
        if not expanded:
@@ -381,22 +370,18 @@ def main(
        if is_batch:
            _run_batch(
                expanded,
-                defaults,
+                request,
                verbose,
                force,
-                ct_explicit,
-                cpu_threads=threads,
                diarize=diarize_enabled,
                speakers=speakers,
            )
        else:
            _run_single(
                expanded[0],
-                defaults,
+                request,
                output,
                verbose,
-                ct_explicit,
-                cpu_threads=threads,
                diarize=diarize_enabled,
                speakers=speakers,
            )
@@ -406,15 +391,15 @@ def main(
    except SystemExit:
        raise
    except ValueError as exc:
-        _print_cuda_hint(exc, resolved_device)
+        _print_cuda_hint(exc, requested_device)
        console.print(f"Ошибка: {exc}", style="red bold")
        raise SystemExit(1)
    except (FileNotFoundError,) as exc:
-        _print_cuda_hint(exc, resolved_device)
+        _print_cuda_hint(exc, requested_device)
        console.print(f"Ошибка: {exc}", style="red bold")
        raise SystemExit(1)
    except Exception as exc:
-        _print_cuda_hint(exc, resolved_device)
+        _print_cuda_hint(exc, requested_device)
        if verbose:
            console.print_exception()
        else:
@@ -423,13 +408,48 @@ def main(
        raise SystemExit(1)


+def _print_execution_header(info: ExecutionInfo, verbose: bool) -> None:
+    """Печатает выбранное исполнение; в --verbose — версии runtime для диагностики."""
+    console.print(
+        f"Модель: [bold]{info.model}[/bold]  "
+        f"Устройство: [bold]{info.device}[/bold]  "
+        f"Compute: [bold]{info.compute_type}[/bold]"
+    )
+    if info.device == "openvino-gpu" and info.model != "large-v3":
+        console.print(
+            "Совет: --model large-v3 даёт лучшее качество на GPU (~2x дольше)",
+            style="dim",
+        )
+    if info.device != info.resolved_device:
+        console.print(
+            f"Запрошено {info.requested_device}, используется {info.device}",
+            style="yellow",
+        )
+    if verbose:
+        threads = info.cpu_threads or "по умолчанию библиотеки"
+        console.print(f"Движок: {info.engine}  Потоки: {threads}", style="dim")
+        for key, value in info.runtime.items():
+            console.print(f"  {key}: {value}", style="dim", markup=False)
+
+
+def _load_diarizer(
+    diarize: bool, speakers: int | None, cpu_threads: int
+) -> SpeakerDiarizer | None:
+    """Готовит диаризатор до первого ASR; пословный контракт уже проверен module."""
+    if not diarize:
+        return None
+    return load_speaker_diarizer(
+        speakers=speakers,
+        threads=cpu_threads,
+        on_status=lambda message: console.print(message),
+    )
+
+
 def _run_single(
    file: Path,
-    defaults: ResolvedConfig,
+    request: ExecutionRequest,
    output: Path | None,
    verbose: bool,
-    compute_type_explicit: bool = False,
-    cpu_threads: int = 0,
    diarize: bool = False,
    speakers: int | None = None,
 ) -> None:
@@ -437,9 +457,6 @@ def _run_single(
    start = time.monotonic()

    validated_file = validate_input_file(file)
-    requested_device = defaults["device"]
-    resolved_device = detect_device(requested_device)
-    strict = requested_device != "auto"
    output_path = build_output_path(validated_file, output)

    console.print(f"Файл: [bold]{validated_file.name}[/bold]")
@@ -447,59 +464,18 @@ def _run_single(
    def on_segment(seg: Segment) -> None:
        console.print(f"  [{seg.start:.2f}s] {seg.text.strip()}")

-    model_obj, actual_device, backend, model_path = load_model(
-        defaults["model"],
-        resolved_device,
-        defaults["compute_type"],
-        on_status=lambda msg: console.print(msg),
-        strict_device=strict,
-        compute_type_explicit=compute_type_explicit,
-        cpu_threads=cpu_threads,
-    )
-    actual_ct = (
-        getattr(backend, "actual_compute_type", defaults["compute_type"])
-        or defaults["compute_type"]
-    )
-    console.print(
-        f"Модель: [bold]{defaults['model']}[/bold]  "
-        f"Устройство: [bold]{actual_device}[/bold]  "
-        f"Compute: [bold]{actual_ct}[/bold]"
-    )
-    if actual_device == "openvino-gpu" and defaults["model"] != "large-v3":
-        console.print(
-            "Совет: --model large-v3 даёт лучшее качество на GPU (~2x дольше)",
-            style="dim",
-        )
-
-    speaker_diarizer = None
-    if diarize:
-        if not backend.word_timestamps_available:
-            raise ValueError(
-                "Выбранный движок или модель не поддерживает пословные таймкоды"
-            )
-        speaker_diarizer = load_speaker_diarizer(
-            speakers=speakers,
-            threads=cpu_threads,
-            on_status=lambda message: console.print(message),
-        )
+    transcriber = Transcriber(request)
+    info = transcriber.prepare(on_status=lambda msg: console.print(msg))
+    _print_execution_header(info, verbose)
+    speaker_diarizer = _load_diarizer(diarize, speakers, request.cpu_threads)

    with Status("Подготавливаю запуск...", console=console) as status:
-        tfr = _transcribe_file(
-            model=model_obj,
-            actual_device=actual_device,
-            backend=backend,
-            model_path=model_path,
-            file_path=validated_file,
-            model_name=defaults["model"],
-            compute_type=defaults["compute_type"],
-            language=defaults["language"] if defaults["language"] != "auto" else None,
+        result = transcriber.transcribe(
+            validated_file,
            on_segment=on_segment if verbose else None,
            on_status=status.update,
-            strict_device=strict,
-            cpu_threads=cpu_threads,
        )

-    result = tfr.result
    speaker_transcript = None
    diarization_warning = None
    diarization_degraded = False
@@ -526,18 +502,12 @@ def _run_single(
        if diarization_warning is not None:
            console.print(f"Внимание: {diarization_warning}", style="yellow")

-    if tfr.actual_device != resolved_device:
-        if requested_device == "auto":
-            console.print(
-                f"Определено устройство {resolved_device}, "
-                f"но использовано {tfr.actual_device} (fallback)",
-                style="yellow",
-            )
-        else:
-            console.print(
-                f"Запрошено {requested_device}, использовано {tfr.actual_device}",
-                style="yellow",
-            )
+    execution = transcriber.execution
+    if execution.device != info.device:
+        console.print(
+            f"Запрошено {execution.requested_device}, использовано {execution.device}",
+            style="yellow",
+        )

    if len(result.segments) == 0:
        message = f"Речь не обнаружена в файле {validated_file.name}"
@@ -545,14 +515,13 @@ def _run_single(
            message += "; диаризация не запускалась"
        console.print(message, style="yellow")

-    device_info = _format_device_info(result.device_used)
-    language_mode = _format_language_mode(defaults["language"], result)
+    language_mode = _format_language_mode(request.language or "auto", result)

    content = format_transcript(
        result=result,
        source_filename=validated_file.name,
-        model_name=defaults["model"],
-        device_info=device_info,
+        model_name=execution.model,
+        device_info=execution.description,
        language_mode=language_mode,
        speaker_transcript=speaker_transcript,
        diarization_warning=diarization_warning,
@@ -569,11 +538,9 @@ def _run_single(

 def _run_batch(
    files: list[Path],
-    defaults: ResolvedConfig,
+    request: ExecutionRequest,
    verbose: bool,
    force: bool,
-    compute_type_explicit: bool = False,
-    cpu_threads: int = 0,
    diarize: bool = False,
    speakers: int | None = None,
 ) -> None:
@@ -603,50 +570,11 @@ def _run_batch(
            raise SystemExit(1)
        return

-    # Phase 2: Load model (ensure + create в одном вызове)
-    requested_device = defaults["device"]
-    resolved_device = detect_device(requested_device)
-    strict = requested_device != "auto"
-    model_obj, actual_device, backend, model_path = load_model(
-        defaults["model"],
-        resolved_device,
-        defaults["compute_type"],
-        on_status=lambda msg: console.print(msg),
-        strict_device=strict,
-        compute_type_explicit=compute_type_explicit,
-        cpu_threads=cpu_threads,
-    )
-
-    if actual_device == "openvino-gpu" and defaults["model"] != "large-v3":
-        console.print(
-            "Совет: --model large-v3 даёт лучшее качество на GPU (~2x дольше)",
-            style="dim",
-        )
-
-    if actual_device != resolved_device:
-        if requested_device == "auto":
-            console.print(
-                f"Определено устройство {resolved_device}, "
-                f"но используется {actual_device} (fallback)",
-                style="yellow",
-            )
-        else:
-            console.print(
-                f"Запрошено {requested_device}, используется {actual_device}",
-                style="yellow",
-            )
-
-    speaker_diarizer = None
-    if diarize:
-        if not backend.word_timestamps_available:
-            raise ValueError(
-                "Выбранный движок или модель не поддерживает пословные таймкоды"
-            )
-        speaker_diarizer = load_speaker_diarizer(
-            speakers=speakers,
-            threads=cpu_threads,
-            on_status=lambda message: console.print(message),
-        )
+    # Phase 2: module выполнения создаётся, когда есть работа
+    transcriber = Transcriber(request)
+    info = transcriber.prepare(on_status=lambda msg: console.print(msg))
+    _print_execution_header(info, verbose)
+    speaker_diarizer = _load_diarizer(diarize, speakers, request.cpu_threads)

    # Phase 3: Transcribe
    processed = 0
@@ -663,39 +591,24 @@ def _run_batch(
            def on_segment(seg: Segment) -> None:
                console.print(f"  [{seg.start:.2f}s] {seg.text.strip()}")

+            device_before = transcriber.execution.device
            with Status(f"{prefix}...", console=console) as status:
-                tfr = _transcribe_file(
-                    model=model_obj,
-                    actual_device=actual_device,
-                    backend=backend,
-                    model_path=model_path,
-                    file_path=file,
-                    model_name=defaults["model"],
-                    compute_type=defaults["compute_type"],
-                    language=defaults["language"]
-                    if defaults["language"] != "auto"
-                    else None,
+                result = transcriber.transcribe(
+                    file,
                    on_segment=on_segment if verbose else None,
                    on_status=status.update
                    if not verbose
                    else lambda msg: console.print(msg),
-                    strict_device=strict,
-                    cpu_threads=cpu_threads,
                )

-            if tfr.actual_device != actual_device:
+            execution = transcriber.execution
+            if execution.device != device_before:
                console.print(
-                    f"  {file.name}: fallback на {tfr.actual_device} при транскрипции",
+                    f"  {file.name}: fallback на {execution.device} при транскрипции",
                    style="yellow",
                )
-            # Обновляем после возможного mid-stream fallback
-            model_obj = tfr.model
-            actual_device = tfr.actual_device
-            backend = tfr.backend
-            model_path = tfr.model_path

-            result = tfr.result
-            language_mode = _format_language_mode(defaults["language"], result)
+            language_mode = _format_language_mode(request.language or "auto", result)
            speaker_transcript = None
            diarization_warning = None
            file_degraded = False
@@ -735,13 +648,11 @@ def _run_batch(
                    message += "; диаризация не запускалась"
                console.print(message, style="yellow")

-            device_info = _format_device_info(result.device_used)
-
            content = format_transcript(
                result=result,
                source_filename=file.name,
-                model_name=defaults["model"],
-                device_info=device_info,
+                model_name=execution.model,
+                device_info=execution.description,
                language_mode=language_mode,
                speaker_transcript=speaker_transcript,
                diarization_warning=diarization_warning,
@@ -760,7 +671,7 @@ def _run_batch(
        except KeyboardInterrupt:
            raise
        except Exception as exc:
-            _print_cuda_hint(exc, actual_device)
+            _print_cuda_hint(exc, request.device)
            if verbose:
                console.print_exception()
            else:
@@ -171,23 +171,3 @@ def _resolve_value(cli_value: _T | None, config_value: _T | None, default: _T) -
    if config_value is not None:
        return config_value
    return default
-
-
-def apply_device_defaults(
-    defaults: ResolvedConfig,
-    resolved_device: str,
-    cli_values: CliValues,
-    config: ConfigValues,
-) -> ResolvedConfig:
-    """Применяет device-aware дефолты для model и compute_type,
-    если они не были явно заданы через CLI или конфиг."""
-    device_defs = DEVICE_DEFAULTS.get(resolved_device, {})
-    if not device_defs:
-        return defaults
-
-    result = dict(defaults)
-    for key in ("model", "compute_type"):
-        if cli_values.get(key) is None and key not in config:
-            if key in device_defs:
-                result[key] = device_defs[key]
-    return cast(ResolvedConfig, result)
@@ -10,9 +10,8 @@ from local_transcriber.backends import get_backend
 from local_transcriber.config import DEVICE_DEFAULTS, HARDCODED_DEFAULTS

 # Re-export из types.py для обратной совместимости
-from local_transcriber.types import (
+from local_transcriber.types import (  # noqa: F401
    Segment,
-    TranscribeFileResult,
    TranscribeResult,
    WordTimestampsUnavailableError,
 )
@@ -228,120 +227,6 @@ def _describe_device(device: str) -> str:
    return "CPU"


-def load_model(
-    model_name: str,
-    device: str,
-    compute_type: str,
-    on_status: Callable[[str], None] | None = None,
-    strict_device: bool = False,
-    compute_type_explicit: bool = False,
-    cpu_threads: int = 0,
-) -> tuple[Any, str, Any, str]:
-    """Загружает модель: ensure + create с fallback.
-
-    Возвращает (model, actual_device, backend, model_path).
-    compute_type_explicit: True если пользователь явно указал --compute-type.
-    cpu_threads: число потоков для CPU inference (0 = дефолт библиотеки).
-    """
-    backend = get_backend(device, compute_type_explicit=compute_type_explicit)
-    actual_device = device
-
-    model_path = backend.ensure_model_available(model_name, compute_type, on_status)
-
-    try:
-        _notify_status(on_status, f"Инициализирую модель на {device}...")
-        model = backend.create_model(
-            model_path, device, compute_type, cpu_threads=cpu_threads
-        )
-        # Резолвим actual_device по реальному OpenVINO device
-        ov_dev = getattr(backend, "actual_ov_device", None)
-        if ov_dev == "GPU" and actual_device != "openvino-gpu":
-            actual_device = "openvino-gpu"
-        elif (
-            ov_dev == "CPU"
-            and actual_device.startswith("openvino")
-            and actual_device != "openvino-cpu"
-        ):
-            actual_device = "openvino-cpu"
-    except (RuntimeError, ValueError) as exc:
-        if device != "cpu" and _is_backend_error(exc, device):
-            if strict_device:
-                raise
-            warnings.warn(
-                f"Не удалось загрузить модель на {device}: {exc}. Переключение на CPU.",
-                stacklevel=2,
-            )
-            actual_device = "cpu"
-            backend = get_backend("cpu")
-            model_path = backend.ensure_model_available(
-                model_name, compute_type, on_status
-            )
-            _notify_status(on_status, "Инициализирую модель на cpu...")
-            model = backend.create_model(
-                model_path, "cpu", compute_type, cpu_threads=cpu_threads
-            )
-        else:
-            raise
-
-    return model, actual_device, backend, model_path
-
-
-def _transcribe_file(
-    model: Any,
-    actual_device: str,
-    backend: Any,
-    model_path: str,
-    file_path: Path,
-    model_name: str,
-    compute_type: str,
-    language: str | None = None,
-    on_segment: Callable[[Segment], None] | None = None,
-    on_status: Callable[[str], None] | None = None,
-    strict_device: bool = False,
-    cpu_threads: int = 0,
-) -> TranscribeFileResult:
-    """Транскрибирует один файл. При mid-stream fallback перезагружает модель."""
-    lang_arg = language if language and language != "auto" else None
-
-    try:
-        _notify_status(on_status, "Транскрибирую...")
-        result = backend.transcribe(model, file_path, lang_arg, on_segment, on_status)
-        result.device_used = actual_device
-    except (RuntimeError, ValueError) as exc:
-        if actual_device != "cpu" and _is_backend_error(exc, actual_device):
-            if strict_device:
-                raise
-            warnings.warn(
-                f"Ошибка при транскрипции на {actual_device}: {exc}. "
-                "Переключение на CPU и повтор.",
-                stacklevel=2,
-            )
-            actual_device = "cpu"
-            backend = get_backend("cpu")
-            model_path = backend.ensure_model_available(
-                model_name, compute_type, on_status
-            )
-            _notify_status(on_status, "Инициализирую модель на cpu...")
-            model = backend.create_model(
-                model_path, "cpu", compute_type, cpu_threads=cpu_threads
-            )
-            _notify_status(on_status, "Транскрибирую...")
-            result = backend.transcribe(
-                model, file_path, lang_arg, on_segment, on_status
-            )
-            result.device_used = actual_device
-        else:
-            raise
-
-    return TranscribeFileResult(
-        result=result,
-        model=model,
-        actual_device=actual_device,
-        backend=backend,
-        model_path=model_path,
-    )
-
-
 def transcribe(
    file_path: Path,
    model_name: str | None = None,
@@ -377,8 +262,6 @@ def ensure_model_available(
    on_status: Callable[[str], None] | None = None,
 ) -> str:
    """Публичный helper: гарантирует наличие модели для указанного бэкенда."""
-    from local_transcriber.config import DEVICE_DEFAULTS, HARDCODED_DEFAULTS
-
    if compute_type is None:
        device_defs = DEVICE_DEFAULTS.get(device, {})
        compute_type = device_defs.get(
@@ -2,7 +2,6 @@

 from collections.abc import Callable
 from dataclasses import dataclass, field
-from typing import Any

 # Единый признак «язык неизвестен» для всех бэкендов
 UNKNOWN_LANGUAGE = "unknown"
@@ -83,14 +82,5 @@ class TranscribeResult:
    words: list[Word] = field(default_factory=list)


-@dataclass
-class TranscribeFileResult:
-    result: TranscribeResult
-    model: Any  # backend-specific model handle
-    actual_device: str
-    backend: Any = None  # backend instance (для переиспользования в батче)
-    model_path: str = ""  # путь к модели (меняется при cross-backend fallback)
-
-
 StatusCallback = Callable[[str], None] | None
 SegmentCallback = Callable[[Segment], None] | None