alignment_subprocess.py

#!/usr/bin/env python3
# alignment_subprocess.py - Subprocess para alineación de WhisperX
import os
import sys
import json
import time
import gc
import torch
import torchaudio
import whisperx
import tempfile
from pathlib import Path

def _inject_local_ffmpeg():
    """Ensure bundled ffmpeg bin folder is in PATH for audio loading in subprocess.
    Safe no-op if not found."""
    try:
        base_dir = Path(__file__).parent
        candidates = [
            base_dir / 'ffmpeg' / 'ffmpeg-8.0-essentials_build' / 'bin',
            base_dir / 'ffmpeg' / 'bin',
        ]
        for c in candidates:
            if c.is_dir():
                bin_path = str(c)
                if bin_path not in os.environ.get('PATH', ''):
                    os.environ['PATH'] = bin_path + os.pathsep + os.environ.get('PATH', '')
                    print(f"🔧 FFmpeg agregado al PATH (align): {bin_path}")
                for exe in ('ffmpeg.exe','ffprobe.exe'):
                    if (c / exe).is_file():
                        print(f"✅ Detectado {exe} en {c}")
                break
        else:
            print("⚠️ FFmpeg local no encontrado para alineación; se usará PATH del sistema")
    except Exception as e:
        print(f"⚠️ Error inyectando FFmpeg (align): {e}")

_inject_local_ffmpeg()

def _get_audio_duration_seconds(path: str) -> float:
    """Try to get duration (seconds) using torchaudio.info; fallback to ffprobe if present.
    Returns 0.0 on failure."""
    try:
        info = torchaudio.info(path)
        if info.num_frames and info.sample_rate:
            return float(info.num_frames) / float(info.sample_rate)
    except Exception:
        pass
    # Fallback to ffprobe if available
    try:
        from shutil import which
        if which("ffprobe"):
            import subprocess
            cmd = [
                "ffprobe", "-v", "error", "-show_entries", "format=duration",
                "-of", "default=noprint_wrappers=1:nokey=1", path
            ]
            out = subprocess.check_output(cmd, text=True).strip()
            return float(out)
    except Exception:
        pass
    return 0.0


def align_segments_subprocess(result_data, audio_path, device="cuda"):
    """
    Realiza alineación de segmentos en subprocess separado para evitar deadlock CUDA
    """
    print("🚀 Iniciando alineación en subprocess...")
    
    try:
        # Verificar que tenemos los datos necesarios
        if not result_data or not result_data.get("segments"):
            print("❌ No hay segmentos para alinear")
            return False

        language = result_data.get("language", "es")
        segments = result_data.get("segments", [])

        print(f"📐 Preparando alineación para {len(segments)} segmentos en idioma {language}")
        
        # Detectar device disponible
        if device == "cuda" and not torch.cuda.is_available():
            device = "cpu"
            print("⚠️ CUDA no disponible, usando CPU para alineación")
        
        if device == "cuda" and torch.cuda.is_available():
            torch.cuda.empty_cache()
            gc.collect()
            free_mem = torch.cuda.mem_get_info()[0] / 1024**3
            print(f"📊 Memoria CUDA inicial: {free_mem:.1f}GB libres")
        
        # Medir duración del audio de trabajo y max end de segmentos
        work_audio_path = str(audio_path)
        orig_audio_path = os.environ.get("ORIGINAL_AUDIO_PATH", "")
        seg_max_end = 0.0
        try:
            seg_max_end = max([(s.get("end") or 0.0) for s in segments]) if segments else 0.0
        except Exception:
            seg_max_end = 0.0
        work_duration = _get_audio_duration_seconds(work_audio_path)
        print(f"📏 Duración audio trabajo: {work_duration:.2f}s | Máx fin segmentos: {seg_max_end:.2f}s")

        # Cargar audio (robusto a diferentes firmas de retorno)
        print(f"📁 Cargando audio: {work_audio_path}")
        try:
            audio_ret = whisperx.load_audio(work_audio_path)
            if isinstance(audio_ret, (list, tuple)):
                audio = audio_ret[0]
            else:
                audio = audio_ret
        except Exception as e:
            raise RuntimeError(f"Error cargando audio con whisperx.load_audio: {e}")
        
        # 🧹 Limpieza antes de cargar modelo de alineación
        if device == "cuda":
            print("🧹 Limpiando memoria antes de cargar modelo de alineación...")
            for _ in range(3):
                torch.cuda.empty_cache()
                torch.cuda.ipc_collect()
                gc.collect()
                time.sleep(0.1)
        
        # Cargar modelo de alineación
        print(f"📐 Cargando modelo de alineación para {language} en {device}...")
        model_a, metadata = whisperx.load_align_model(language_code=language, device=device)
        
        if device == "cuda":
            allocated = torch.cuda.memory_allocated(0) / 1024**3
            print(f"📊 Memoria tras cargar modelo: {allocated:.1f}GB asignada")
        
        # Decidir si reintentar con original en caso de duración insuficiente
        retried_with_original = False
        audio_path_used = work_audio_path

        def _run_align(audio_array):
            print(f"🎯 Alineando {len(segments)} segmentos...")
            return whisperx.align(segments, model_a, metadata, audio_array, device)

        try:
            # Pre-chequeo: si duración de trabajo parece menor que los segmentos
            if work_duration and seg_max_end and (work_duration + 0.25 < seg_max_end) and orig_audio_path and os.path.isfile(orig_audio_path):
                print("⚠️ Audio de trabajo parece más corto que los segmentos. Reintentando con audio ORIGINAL...")
                try:
                    audio_ret2 = whisperx.load_audio(orig_audio_path)
                    audio2 = audio_ret2[0] if isinstance(audio_ret2, (list, tuple)) else audio_ret2
                    aligned = _run_align(audio2)
                    retried_with_original = True
                    audio_path_used = orig_audio_path
                except Exception as e2:
                    print(f"⚠️ Falló alineación con original en pre-chequeo: {e2}. Probando audio de trabajo igualmente...")
                    aligned = _run_align(audio)
            else:
                aligned = _run_align(audio)
        except Exception as e:
            # Error durante alineación: si no hemos probado original y existe, reintentar una vez
            if (not retried_with_original) and orig_audio_path and os.path.isfile(orig_audio_path):
                print(f"⚠️ Error en alineación inicial: {e}. Reintentando con audio ORIGINAL...")
                try:
                    audio_ret2 = whisperx.load_audio(orig_audio_path)
                    audio2 = audio_ret2[0] if isinstance(audio_ret2, (list, tuple)) else audio_ret2
                    aligned = _run_align(audio2)
                    retried_with_original = True
                    audio_path_used = orig_audio_path
                except Exception as e2:
                    # Añadir contexto y relanzar si sigue fallando
                    raise RuntimeError(f"Error durante whisperx.align (reintento con original también falló): {e}; reintento: {e2}")
            else:
                # Añadir contexto y relanzar para el handler exterior
                raise RuntimeError(f"Error durante whisperx.align: {e}")
        
        # 🧹 Limpiar modelo de memoria
        print("🧹 Liberando modelo de alineación...")
        del model_a
        del metadata
        try:
            del audio
        except Exception:
            pass
        
        if device == "cuda":
            for _ in range(3):
                torch.cuda.empty_cache()
                torch.cuda.ipc_collect()
                gc.collect()
                time.sleep(0.1)
            
            final_allocated = torch.cuda.memory_allocated(0) / 1024**3
            print(f"📊 Memoria final: {final_allocated:.1f}GB asignada")
        
        # Guardar resultado
        aligned_segments = aligned.get("segments", [])
        result_data_aligned = {
            "success": True,
            "result": {
                "segments": aligned_segments,
                "language": language
            },
            "aligned_count": len(aligned_segments),
            "device": device,
            "audio_path": str(audio_path),
            "audio_path_used": audio_path_used,
            "retried_with_original": retried_with_original,
            "audio_duration_sec": work_duration,
            "segments_max_end_sec": seg_max_end
        }
        
        with open("alignment_result.json", "w", encoding="utf-8") as f:
            json.dump(result_data_aligned, f, ensure_ascii=False, indent=2)
            
        print(f"✅ Alineación completada: {len(aligned_segments)} segmentos alineados")
        return True
        
    except Exception as e:
        # 🧹 Limpiar memoria en caso de error
        if device == "cuda" and torch.cuda.is_available():
            torch.cuda.empty_cache()
            gc.collect()

        import traceback
        tb = traceback.format_exc()

        # Guardar error con traceback completo para diagnóstico
        error_data = {
            "success": False,
            "error": str(e),
            "traceback": tb,
            "audio_path": str(audio_path)
        }

        try:
            with open("alignment_result.json", "w", encoding="utf-8") as f:
                json.dump(error_data, f, ensure_ascii=False, indent=2)
        except Exception as save_err:
            print(f"⚠️ No se pudo escribir alignment_result.json: {save_err}")

        print(f"❌ Error en alineación: {e}")
        print(f"❌ Traceback: {tb}")
        return False

if __name__ == "__main__":
    try:
        if len(sys.argv) < 3:
            print("❌ Error: Faltan argumentos")
            print("Uso: python alignment_subprocess.py <transcription_result.json> <audio_path> [device]")
            sys.exit(1)
        
        transcription_file = sys.argv[1]
        audio_path = sys.argv[2]
        device = sys.argv[3] if len(sys.argv) > 3 else "cuda"
        
        print(f"🚀 Iniciando alineación subprocess: {audio_path}")
        print(f"📄 Archivo transcripción: {transcription_file}")
        print(f"🖥️ Device: {device}")
        
        # Cargar datos de transcripción
        with open(transcription_file, "r", encoding="utf-8") as f:
            result_data = json.load(f)
        
        if not result_data.get("success"):
            print("❌ Los datos de transcripción no son válidos")
            sys.exit(1)
            
        transcription_result = result_data.get("result", {})
        
        # Realizar alineación
        success = align_segments_subprocess(transcription_result, audio_path, device)
        
        if success:
            print("✅ Alineación subprocess completada exitosamente")
            sys.exit(0)
        else:
            print("❌ Alineación subprocess falló")
            sys.exit(1)
            
    except Exception as e:
        print(f"❌ Error crítico en alignment subprocess: {e}")
        import traceback
        print(f"❌ Traceback: {traceback.format_exc()}")
        
        # Guardar error para el proceso principal
        error_data = {
            "success": False,
            "error": f"Error crítico: {str(e)}",
            "traceback": traceback.format_exc()
        }
        
        try:
            with open("alignment_result.json", "w", encoding="utf-8") as f:
                json.dump(error_data, f, ensure_ascii=False, indent=2)
        except Exception as save_error:
            print(f"❌ No se pudo guardar error: {save_error}")
        
        sys.exit(1)