Fix model dependencies and warnings

- Add protobuf>=3.20.0 to requirements.txt (required for DeepSeek tokenizer)
- Add warning suppression for FutureWarning and rope_scaling messages
- Suppress slow image processor warnings
- These changes address the tokenizer loading issues with the new DeepSeek model

backend_service.py

@@ -4,6 +4,13 @@ Provides OpenAI-compatible chat completion endpoints
 """
 
 import os
+import warnings
+
+# Suppress warnings before any other imports
+warnings.filterwarnings("ignore", category=FutureWarning, module="transformers")
+warnings.filterwarnings("ignore", message=".*slow image processor.*")
+warnings.filterwarnings("ignore", message=".*rope_scaling.*")
+
 # Direct Hugging Face caches to a writable folder under /tmp (use only HF_HOME, TRANSFORMERS_CACHE is deprecated)
 os.environ.setdefault("HF_HOME", "/tmp/.cache/huggingface")
 # Suppress advisory warnings from transformers (including deprecation warnings)

requirements.txt

@@ -5,6 +5,7 @@ torch>=2.0.0
 Pillow>=10.0.0
 accelerate>=0.24.0
 requests>=2.31.0
+protobuf>=3.20.0
 # NOTE: GGUF models like 'unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF' must be downloaded manually or referenced from HuggingFace, not pip-installed.
 fastapi>=0.100.0
 uvicorn[standard]>=0.23.0