models/cpp_qwen2.py

"""
https://github.com/abetlen/llama-cpp-python/blob/main/examples/gradio_chat/local.py
https://github.com/awinml/llama-cpp-python-bindings

python convert_hf_to_gguf.py --outtype f16 Qwen1.5-0.5B-Chat

python convert_hf_to_gguf.py /workspace/xusong/huggingface/models/Qwen2-0.5B-Instruct/


./llama-cli -m /workspace/xusong/huggingface/models/Qwen1.5-0.5B-Chat/Qwen1.5-0.5B-Chat-F16.gguf -p "I believe the meaning of life is" -n 128

./llama-cli -m /workspace/xusong/huggingface/models/Qwen1.5-0.5B-Chat/Qwen1.5-0.5B-Chat-F16.gguf -f prompt.txt -n 128

./llama-cli -m /workspace/xusong/huggingface/models/Qwen1.5-0.5B-Chat/Qwen1.5-0.5B-Chat-F16.gguf -p "You are a helpful assistant" -cnv


## reference

- https://github.com/langchain-ai/langchain/blob/master/libs/community/langchain_community/llms/llamacpp.py
- https://github.com/abetlen/llama-cpp-python/blob/main/examples/gradio_chat/server.py
- https://github.com/abetlen/llama-cpp-python/blob/main/llama_cpp/server/app.py

"""

import json
import copy
import os

from models.base_model import Simulator
import llama_cpp
# import llama_cpp.llama_tokenizer
from transformers import AutoTokenizer
from utils.logging_util import logger
import config


class Qwen2Simulator(Simulator):

    def __init__(self):
        local_path = "/workspace/xusong/huggingface/models/Qwen2-0.5B-Instruct-GGUF/qwen2-0_5b-instruct-fp16.gguf"
        if os.path.exists(local_path):
            self.hf_tokenizer = AutoTokenizer.from_pretrained(
                "/workspace/xusong/huggingface/models/Qwen2-0.5B-Instruct/")
            self.llm = llama_cpp.Llama(  # n_ctx, n_threads
                model_path=local_path,
                # 默认的tokenizer有bug，tokenize后的id不同
                tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer(self.hf_tokenizer),
                n_ctx=config.MAX_SEQUENCE_LENGTH,  #
                # n_threads=None, # 默认会根据cpu数来设置 n_threads
                # use_mlock=True,
                verbose=True,
            )
        else:
            self.hf_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
            self.llm = llama_cpp.Llama.from_pretrained(
                repo_id="Qwen/Qwen2-0.5B-Instruct-GGUF",
                tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer(self.hf_tokenizer),
                filename="*fp16.gguf",
                n_ctx=config.MAX_SEQUENCE_LENGTH,
                # use_mlock=True,
                verbose=False,
            )
        logger.info(f"llm has been initialized: {self.llm}, "
                    f"n_threads={self.llm.n_threads}, n_ctx={self.llm.n_ctx}, "
                    f"env[CACHE]={os.environ.get('CACHE', None)}")

        self.generation_kwargs = dict(
            temperature=config.DEFAULT_TEMPERATURE,
            top_p=config.DEFAULT_TOP_P,
            top_k=config.DEFAULT_TOP_K,
            max_tokens=config.DEFAULT_MAX_TOKENS,
            repeat_penalty=1.1,
            # qwen2-0.5b-chat 有时内容生成结束没有<|im_end|>，直接跟 <|im_start|>
            stop=[
                "<|im_end|>",
                "<|im_start|>",
                "<|endoftext|>",
            ],
        )

    def tokenize(self, text):
        return self.llm.tokenize(text.encode("utf-8"))


    def generate(self, history, stream=True):
        if history[-1]['role'] in ["user"]:
            start_tokens = self.tokenize("<|im_start|>assistant\n")
        elif history[-1]['role'] in ["assistant", "system"]:
            start_tokens = self.tokenize("<|im_start|>user\n")

        input_ids = []
        for message in history:
            if "tokens" not in message:
                message["tokens"] = self.tokenize(message["content"])
            input_ids += self.tokenize(f"<|im_start|>{message['role']}\n") \
                         + message["tokens"] \
                         + self.tokenize("<|im_end|>\n")
        input_ids += start_tokens
        if stream:
            return self._stream_generate(input_ids)
        else:
            return self._generate(input_ids)

    def _stream_generate(self, input_ids):
        logger.info(f"generation_kwargs {self.generation_kwargs}")

        # self.llm.generate  .set_cache   .last_n_tokens_size  .reset  .ctx ._ctx
        output = self.llm.create_completion(
            input_ids,
            stream=True,
            **self.generation_kwargs
        )
        # TODO: 检测finish reason，如果是length，则shift，并继续生成。
        # TODO: 返回 token_id,
        for out in output:
            stream = copy.deepcopy(out)
            if stream["choices"][0]["finish_reason"] is None:
                yield stream["choices"][0]["completion_text"], stream["choices"][0]["completion_tokens"]


bot = Qwen2Simulator()

if __name__ == "__main__":

    messages = [{"role": "system", "content": "你是一个导游。"}]
    generated_tokens = None
    print("######## requesting", messages)
    for generated_text, generated_tokens in bot.generate(messages, stream=True):
        print(generated_text, generated_tokens)


    for i in range(3):
        messages.append({"role": "user" if i % 2 == 0 else "assistant", "content": generated_text, "tokens": generated_tokens})
        print("######## requesting", messages)
        for generated_text, generated_tokens in bot.generate(messages, stream=True):
            pass
            # print(generated_text, all_tokens)