Fix double task_done in ASR/detect-language handlers (#262)

Add a check to ensure model exists before unloading.

.github/workflows/build_GPU.yml

@@ -3,17 +3,17 @@ name: Build_Subgen_Dockerfile_GPU
 on:
   push:
     paths:
-      - requirements.txt
-      - Dockerfile
+      - 'requirements.txt'
+      - 'Dockerfile'
   workflow_dispatch:
 
 jobs:
   docker:
-    runs-on: ubuntu-latest
+    runs-on: [self-hosted]
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v2
         with:
           fetch-depth: 0
 
@@ -21,27 +21,7 @@ jobs:
         id: get_version
         run: |
           version=$(grep -oP "subgen_version\s*=\s*'\K[^']+" subgen.py)
-          echo "version=$version" >> "$GITHUB_ENV"
-          
-      - name: Pre-job:free disk space
-        shell: bash
-        run: |
-          set -euo pipefail
-          echo "Disk before cleanup:"
-          df -h /
-
-          # Prune buildx builder data and docker caches
-          sudo docker builder prune --all --force || true
-          sudo docker system prune -af --volumes || true
-
-          # Remove temp files
-          sudo rm -rf /tmp/* || true
-
-          # Remove old buildkit cache if present
-          sudo rm -rf /var/lib/docker/buildkit || true
-
-          echo "Disk after cleanup:"
-          df -h /
+          echo "version=$version" >> $GITHUB_ENV
 
       - name: Login to Docker Hub
         uses: docker/login-action@v3
@@ -57,14 +37,3 @@ jobs:
           tags: |
             mccloud/subgen:latest
             mccloud/subgen:${{ env.version }}
-
-      - name: Post-job:cleanup local docker artifacts
-        if: always()
-        shell: bash
-        run: |
-          set -euo pipefail
-          sudo docker builder prune --all --force || true
-          sudo docker image prune -af || true
-          sudo docker container prune -f || true
-          sudo docker system prune -af --volumes || true
-          df -h /

Dockerfile

@@ -1,5 +1,5 @@
 # Stage 1: Builder
-FROM nvidia/cuda:12.6.3-base-ubuntu22.04 AS builder
+FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04 AS builder
 
 WORKDIR /subgen
 
@@ -22,7 +22,7 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY . .
 
 # Stage 2: Runtime
-FROM nvidia/cuda:12.6.3-base-ubuntu22.04
+FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
 
 WORKDIR /subgen
 

subgen.py

@@ -1,4 +1,4 @@
-subgen_version = '2025.11.1'
+subgen_version = '2026.01.5'
 
 """
 ENVIRONMENT VARIABLES DOCUMENTATION
@@ -74,6 +74,7 @@ from io import BytesIO
 import io
 import asyncio
 import torch
+import ctypes, ctypes.util
 from typing import List
 from enum import Enum
 
@@ -265,24 +266,24 @@ task_queue = DeduplicatedQueue()
 
 def transcription_worker():
     while True:
-        try:        
+        task = None
+        try:
             task = task_queue.get(block=True, timeout=1)
             if "type" in task and task["type"] == "detect_language":
                 detect_language_task(task['path'])
             elif 'Bazarr-' in task['path']:
                 logging.info(f"Task {task['path']} is being handled by ASR.")
             else:
-                logging.info(f"Task {task['path']} is being handled by Subgen.") 
+                logging.info(f"Task {task['path']} is being handled by Subgen.")
                 gen_subtitles(task['path'], task['transcribe_or_translate'], task['force_language'])
-                task_queue.task_done()
-            # show queue
-            logging.debug(f"Queue status: {task_queue.qsize()} tasks remaining")
         except queue.Empty:
-            continue # This is ok, as we have a timeout, nothing needs to be printed
+            continue
         except Exception as e:
-            logging.error(f"Error processing task: {e}", exc_info=True) # Log the error and the traceback
-        else:
-            delete_model()  # Call delete_model() *only* if no exception occurred
+            logging.error(f"Error processing task: {e}", exc_info=True)
+        finally:
+            if task: # Ensure a task was actually retrieved before calling task_done
+                task_queue.task_done()
+            delete_model()
 
 for _ in range(concurrent_transcriptions):
     threading.Thread(target=transcription_worker, daemon=True).start()
@@ -585,7 +586,6 @@ async def asr(
     
     finally:
         await audio_file.close()
-        task_queue.task_done()
         delete_model()
     
     if result:
@@ -665,7 +665,6 @@ async def detect_language(
         
     finally:
         #await audio_file.close()
-        task_queue.task_done()
         delete_model()
 
         return {"detected_language": detected_language.to_name(), "language_code": language_code}
@@ -796,14 +795,16 @@ def delete_model():
     global model
     if clear_vram_on_complete and task_queue.is_idle():
         logging.debug("Queue idle; clearing model from memory.")
-        model.model.unload_model()
-        del model
-        model = None
+        if model:
+            model.model.unload_model()
+            del model
+            model = None
         if transcribe_device.lower() == 'cuda' and torch.cuda.is_available():
             torch.cuda.empty_cache()
             logging.debug("CUDA cache cleared.")
     if os.name != 'nt': # don't garbage collect on Windows, it will crash the script
         gc.collect()
+        ctypes.CDLL(ctypes.util.find_library('c')).malloc_trim(0)
 
 def isAudioFileExtension(file_extension):
     return file_extension.casefold() in \
@@ -1139,6 +1140,7 @@ def find_language_audio_track(audio_tracks, find_languages):
             if track['language'] == language:
                 return language
     return None
+
 def find_default_audio_track_language(audio_tracks):    
     """
     Finds the language of the default audio track in the given list of audio tracks.
@@ -1255,6 +1257,7 @@ def should_skip_file(file_path: str, target_language: LanguageCode) -> bool:
     logging.debug(f"Processing {base_name}: No skip conditions met.")
     return False
     
+
 def get_subtitle_languages(video_path):
     """
     Extract language codes from each audio stream in the video file using pyav.