app.py

import os
import time

# ruff: noqa: E402
# os.system("pip install --upgrade torch transformers sentencepiece scipy cpm_kernels accelerate bitsandbytes loguru")

os.system("pip install torch transformers sentencepiece loguru")

from pathlib import Path

import torch
from logru import logger
from transformers import AutoModel, AutoTokenizer

# fix timezone in Linux
os.environ["TZ"] = "Asia/Shanghai"
try:
    time.tzset()  # type: ignore # pylint: disable=no-member
except Exception:
    # Windows
    logger.warning("Windows, cant run time.tzset()")

model_name = "THUDM/chatglm2-6b-int4"  # 3.9G

tokenizer = AutoTokenizer.from_pretrained(
    "THUDM/chatglm2-6b-int4", trust_remote_code=True
)

has_cuda = torch.cuda.is_available()
# has_cuda = False  # force cpu

logger.debug("load")
if has_cuda:
    if model_name.endswith("int4"):
        model = AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda()
    else:
        model = (
            AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda().half()
        )
else:
    model = AutoModel.from_pretrained(
        model_name, trust_remote_code=True
    ).half()  # .float() .half().float()

model = model.eval()
logger.debug("done load")

# tokenizer = AutoTokenizer.from_pretrained("openchat/openchat_v2_w")
# model = AutoModelForCausalLM.from_pretrained("openchat/openchat_v2_w", load_in_8bit_fp32_cpu_offload=True, load_in_8bit=True)

model_path = model.config._dict["model_name_or_path"]
logger.debug(f"{model_path=}")

model_size_gb = Path(model_path).stat().st_size / 2**30

logger.info(f"{model_name=} {model_size_gb=:.2f} GB")

# with gr.Blocks() as demo:
#     chatbot = gr.Chatbot()
#     msg = gr.Textbox()
#     clear = gr.ClearButton([msg, chatbot])

#     def respond(message, chat_history):
#         response, chat_history = model.chat(tokenizer, message, history=chat_history, temperature=0.7, repetition_penalty=1.2, max_length=128)
#         chat_history.append((message, response))
#         return "", chat_history

#     msg.submit(respond, [msg, chatbot], [msg, chatbot])

# demo.launch()