import sys
sys.path.append("..")

import json
import torch

from modeling_chatglm import ChatGLMForConditionalGeneration
from tokenization_chatglm import ChatGLMTokenizer
from modelscope import snapshot_download

from tools import show

from transformers import AutoConfig

seed = 4321
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)

pretrained_model_name_or_path = snapshot_download("ZhipuAI/chatglm3-6b")
config, kwargs = AutoConfig.from_pretrained(
    pretrained_model_name_or_path,
    return_unused_kwargs=True,
    trust_remote_code=True,
    code_revision=None,
    _commit_hash=None,
)
glm = ChatGLMForConditionalGeneration(config)


tokenizer_config_file = "./tokenizer_config.json"
if tokenizer_config_file is not None:
    with open(tokenizer_config_file, encoding="utf-8") as tokenizer_config_handle:
        init_kwargs = json.load(tokenizer_config_handle)
    init_kwargs.pop("tokenizer_class", None)
    init_kwargs.pop("tokenizer_file", None)
    saved_init_inputs = init_kwargs.pop("init_inputs", ())
    init_inputs = saved_init_inputs
init_kwargs["vocab_file"] = "./tokenizer.model"
init_kwargs["added_tokens_file"] = None
init_kwargs["special_tokens_map_file"] = None
init_kwargs["tokenizer_file"] = None
init_kwargs["name_or_path"] = pretrained_model_name_or_path
tokenizer = ChatGLMTokenizer(*init_inputs, **init_kwargs)


glm = glm.from_pretrained(pretrained_model_name_or_path).half().cuda()
query = "你好"
response = glm.backward(tokenizer, query)