load the ASR model once per process

- keep WhisperX, align and GigaAM caches process-wide and thread-safe\n- preload the ASR model at the start of the staged stage 3\n- log that the model is loaded once and then reused\n- never clear caches between files, batches or OOM retries

core/pipeline.py

@@ -19,7 +19,9 @@ from .lyrics import LyricsMatch, search_by_identity, search_by_text
 from .metadata import TrackMetadata, read_metadata
 from .recognize import ShazamResult, recognize
 from .separate import separate_vocals, separate_vocals_batch
-from .transcribe import CudaOOMError, transcribe
+from .transcribe import (
+    CudaOOMError, preload_transcription_model, transcribe,
+)
 from .lrc import is_lrc, read_embedded_lyrics
 
 
@@ -1261,6 +1263,20 @@ def _run_staged_batch(
         if config.transcribe.enabled and not cancelled():
             todo = active()
             log(f"\n══ Этап 3 · Транскрибация ({len(todo)} файлов) ══")
+            if todo:
+                # Load once, after separation has finished. Every later file
+                # and every later batch reuses this instance.
+                try:
+                    preload_transcription_model(
+                        engine=config.transcribe.engine,
+                        model=config.transcribe.model,
+                        device=config.transcribe.device,
+                        compute_type=config.transcribe.compute_type,
+                        gigaam_provider=config.transcribe.gigaam_provider,
+                        local_files_only=config.transcribe.local_files_only,
+                    )
+                except Exception as e:  # noqa: BLE001
+                    log(f"  ⚠ Не удалось загрузить модель заранее: {e}")
             progress("transcribe", 0, len(todo))
             for i, f in enumerate(todo, 1):
                 if cancelled():

core/transcribe.py

@@ -6,11 +6,21 @@ Models are cached; aggressive VRAM cleanup happens after each file.
 from __future__ import annotations
 
 import gc
+import threading
 import time
 from pathlib import Path
 from typing import Optional
 
 
+# The model caches below are process-wide. They are intentionally never
+# cleared between files or batches: reloading WhisperX takes minutes and is
+# what made runs look like they reloaded the model every batch. The lock makes
+# the "load once, reuse forever" guarantee safe even if ASR is ever called
+# from multiple threads.
+_MODEL_LOCK = threading.RLock()
+_REUSE_LOGGED: set = set()
+
+
 class CudaOOMError(RuntimeError):
     """CUDA OOM, переживший все стратегии понижения нагрузки."""
 
@@ -110,6 +120,7 @@ class _WhisperXModels:
         if compute_type is None:
             compute_type = "int8_float16" if device == "cuda" else "int8"
         key = (model_name, device, compute_type, local_files_only)
+        with _MODEL_LOCK:
             if key not in cls._cache:
                 if device == "cuda":
                     _enable_tf32()
@@ -125,7 +136,10 @@ class _WhisperXModels:
                     )
 
                 cls._cache[key] = _retry(load)
-            print("  [ASR] model loaded.")
+                print("  [ASR] model loaded once — subsequent files will reuse it.")
+            elif ("whisperx",) + key not in _REUSE_LOGGED:
+                _REUSE_LOGGED.add(("whisperx",) + key)
+                print(f"  [ASR] reusing loaded WhisperX {model_name} ({device}).")
             return cls._cache[key]
 
 
@@ -135,6 +149,7 @@ class _AlignModels:
     @classmethod
     def get(cls, language_code: str, device: str, model_name: Optional[str] = None):
         key = (language_code, device, model_name or "")
+        with _MODEL_LOCK:
             if key not in cls._cache:
                 import whisperx
                 print(f"  [ALIGN] loading align model for {language_code} "
@@ -147,7 +162,10 @@ class _AlignModels:
                     return whisperx.load_align_model(**kwargs)
 
                 cls._cache[key] = _retry(load)
-            print("  [ALIGN] align model loaded.")
+                print("  [ALIGN] align model loaded once — reusing it.")
+            elif ("align",) + key not in _REUSE_LOGGED:
+                _REUSE_LOGGED.add(("align",) + key)
+                print(f"  [ALIGN] reusing loaded align model for {language_code}.")
             return cls._cache[key]
 
 
@@ -162,6 +180,7 @@ class _GigaAMModels:
             "auto": ["CUDAExecutionProvider", "CPUExecutionProvider"],
         }.get(provider, ["CPUExecutionProvider"])
         key = tuple(providers)
+        with _MODEL_LOCK:
             if key not in cls._cache:
                 import onnx_asr
 
@@ -172,6 +191,28 @@ class _GigaAMModels:
             return cls._cache[key]
 
 
+def preload_transcription_model(
+    engine: str = "whisperx",
+    model: str = "large-v3",
+    device: str = "auto",
+    compute_type: Optional[str] = None,
+    gigaam_provider: str = "auto",
+    local_files_only: bool = False,
+) -> None:
+    """Load the ASR model once so the per-file loop only reuses it.
+
+    All model caches are process-wide and are never cleared between files,
+    batches or pipeline runs. This helper just makes the first load explicit.
+    """
+    device = _resolve_device(device)
+    if engine == "whisperx":
+        _WhisperXModels.get(model, device, compute_type, local_files_only)
+    elif engine == "gigaam":
+        _GigaAMModels.get(gigaam_provider)
+    else:
+        raise ValueError(f"Unknown ASR engine: {engine!r}")
+
+
 def transcribe(
     audio_path: Path,
     engine: str = "whisperx",