perf: update dtype to bfloat16

perf: convert weights from float32 to bfloat16

fix: change tokenizer_class from TokenizersBackend to Qwen2Tokenizer for vLLM compat

Upload folder using huggingface_hub

initial commit