Replace model with mpt-7b-instruct, loaded in f16 and sharded to 2GB chunks

increase max_new_tokens default

better generation params

add sharded checkpoint