Patch remote config typo (ministral3 -> mistral3) before loading model

Pin transformers to 4.50.0 for mistral3 support

Bypass TokenizersBackend bug by using Nemo tokenizer for Ministral

Pin transformers to 4.48.2 to avoid broken remote versions

Add offload_folder to support low-memory loading on Hugging Face

Fix Streamlit set_page_config order and print specific ImportError

Final cloud stability fixes for PEFT adapter loading

Fix NoneType error by conditionally passing quantization_config

Fix FP8 quantization conflict by skipping manual config for Mistral-3

Fix TypeError by using explicit BitsAndBytesConfig for quantization

Fix TokenizersBackend error by using slow tokenizer and adding mistral-common

Fix OSError by switching to AutoTokenizer for cloud loading

Optimize cloud build: use Python 3.11 and remove vllm

Fix VLM architecture incompatibility (Ministral/Qwen)

Corrected Hugging Face repository IDs and case

Fix Streamlit popover attribute error by pinning version 1.35.0

Fix Hugging Face Space configuration

Add Cloud compatibility & Transformers fallback

Fix cloud deployment crash and add transformers fallback

Simplified project to local MLX only