feat: Limit concurrent image generations to 4 using an asyncio semaphore.

api/service/generation_service.py

@@ -22,6 +22,9 @@ from adapters.s3_adapter import S3Adapter
 
 logger = logging.getLogger(__name__)
 
+# Limit concurrent generations to 4
+generation_semaphore = asyncio.Semaphore(4)
+
 
 # --- Вспомогательная функция генерации ---
 async def generate_image_task(
@@ -141,8 +144,10 @@ class GenerationService:
             generation_model.id = gen_id
 
             async def runner(gen):
-                logger.info(f"Starting background generation task for ID: {gen.id}")
+                logger.info(f"Generation {gen.id} entered queue (waiting for slot)...")
                 try:
+                    async with generation_semaphore:
+                        logger.info(f"Starting background generation task for ID: {gen.id}")
                         await self.create_generation(gen)
                         logger.info(f"Background generation task finished for ID: {gen.id}")
                 except Exception: