fix: add eoa into eos_token_id in chat to accelerate chat interface

use bin instead of safetensors with max shard of 2GB

fix(modeling): fix inference code

initial commit internlm2-chat-20b model