Add HuggingFace token for gated model access

docker-compose.yml

@@ -39,6 +39,8 @@ services:
       - MAX_MODEL_LEN=2048
       - CACHE_ENABLED=true
       - RETENTION_DAYS=10
+      - HF_TOKEN=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
+      - HUGGING_FACE_HUB_TOKEN=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
     
     # Resource limits (adjust based on your Orin config)
     deploy:

main.py

@@ -39,7 +39,7 @@ from fastapi.responses import FileResponse, StreamingResponse
 from pydantic import BaseModel
 
 # Configuration from environment
-ORPHEUS_MODEL = os.getenv("ORPHEUS_MODEL", "medium-3b")
+ORPHEUS_MODEL = os.getenv("ORPHEUS_MODEL", "canopylabs/orpheus-tts-0.1-finetune-prod")
 CACHE_ENABLED = os.getenv("CACHE_ENABLED", "true").lower() == "true"
 CACHE_DIR = Path(os.getenv("CACHE_DIR", "cache"))
 OUTPUT_DIR = Path(os.getenv("OUTPUT_DIR", "output"))