ai/utils.py

import gc
import json
import os
from typing import Dict, Union, Optional, Tuple

import torch
from torch.nn import Module
from transformers import AutoModel, PreTrainedModel, PreTrainedTokenizer
from transformers.generation.logits_process import LogitsProcessor


def auto_configure_device_map(num_gpus: int) -> Dict[str, int]:
    # transformer.word_embeddings 占用1层
    # transformer.final_layernorm 和 lm_head 占用1层
    # transformer.layers 占用 28 层
    # 总共30层分配到num_gpus张卡上
    num_trans_layers = 28
    per_gpu_layers = 30 / num_gpus

    # bugfix: 在linux中调用torch.embedding传入的weight,input不在同一device上,导致RuntimeError
    # windows下 model.device 会被设置成 transformer.word_embeddings.device
    # linux下 model.device 会被设置成 lm_head.device
    # 在调用chat或者stream_chat时,input_ids会被放到model.device上
    # 如果transformer.word_embeddings.device和model.device不同,则会导致RuntimeError
    # 因此这里将transformer.word_embeddings,transformer.final_layernorm,lm_head都放到第一张卡上
    # 本文件来源于https://github.com/THUDM/ChatGLM-6B/blob/main/utils.py
    # 仅此处做少许修改以支持ChatGLM3
    device_map = {
        'transformer.embedding.word_embeddings': 0,
        'transformer.encoder.final_layernorm': 0,
        'transformer.output_layer': 0,
        'transformer.rotary_pos_emb': 0,
        'lm_head': 0
    }

    used = 2
    gpu_target = 0
    for i in range(num_trans_layers):
        if used >= per_gpu_layers:
            gpu_target += 1
            used = 0
        assert gpu_target < num_gpus
        device_map[f'transformer.encoder.layers.{i}'] = gpu_target
        used += 1

    return device_map


def load_model_on_gpus(checkpoint_path: Union[str, os.PathLike], num_gpus: int = 2,
                       device_map: Optional[Dict[str, int]] = None, **kwargs) -> Module:
    if num_gpus < 2 and device_map is None:
        model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs).half().cuda()
    else:
        from accelerate import dispatch_model

        model = AutoModel.from_pretrained(checkpoint_path, trust_remote_code=True, **kwargs).half()

        if device_map is None:
            device_map = auto_configure_device_map(num_gpus)

        model = dispatch_model(model, device_map=device_map)

    return model


class InvalidScoreLogitsProcessor(LogitsProcessor):
    def __call__(
        self, input_ids: torch.LongTensor, scores: torch.FloatTensor
    ) -> torch.FloatTensor:
        if torch.isnan(scores).any() or torch.isinf(scores).any():
            scores.zero_()
            scores[..., 5] = 5e4
        return scores


def process_response(output: str, use_tool: bool = False) -> Union[str, dict]:
    content = ""
    for response in output.split("<|assistant|>"):
        metadata, content = response.split("\n", maxsplit=1)
        if not metadata.strip():
            content = content.strip()
            content = content.replace("[[训练时间]]", "2023年")
        else:
            if use_tool:
                content = "\n".join(content.split("\n")[1:-1])

                def tool_call(**kwargs):
                    return kwargs

                parameters = eval(content)
                content = {
                    "name": metadata.strip(),
                    "arguments": json.dumps(parameters, ensure_ascii=False)
                }
            else:
                content = {
                    "name": metadata.strip(),
                    "content": content
                }
    return content


@torch.inference_mode()
def generate_stream_chatglm3(model: PreTrainedModel, tokenizer: PreTrainedTokenizer, params: dict):
    messages = params["messages"]
    functions = params["functions"]
    temperature = float(params.get("temperature", 1.0))
    repetition_penalty = float(params.get("repetition_penalty", 1.0))
    top_p = float(params.get("top_p", 1.0))
    max_new_tokens = int(params.get("max_tokens", 256))
    max_length = params.get("max_length", None)
    echo = params.get("echo", True)

    messages = process_chatglm_messages(messages, functions=functions)
    query, role = messages[-1]["content"], messages[-1]["role"]

    inputs = tokenizer.build_chat_input(query, history=messages[:-1], role=role)
    inputs = inputs.to(model.device)
    input_echo_len = len(inputs["input_ids"][0])

    if input_echo_len >= model.config.seq_length:
        print(f"Input length larger than {model.config.seq_length}")

    if max_length is None:
        max_length = min(max_new_tokens + input_echo_len, model.config.seq_length)

    eos_token_id = [
        tokenizer.eos_token_id,
        tokenizer.get_command("<|user|>"),
    ]

    gen_kwargs = {
        "max_length": max_length,
        "do_sample": True if temperature > 1e-5 else False,
        "top_p": top_p,
        "repetition_penalty": repetition_penalty,
        "logits_processor": [InvalidScoreLogitsProcessor()],
    }
    if temperature > 1e-5:
        gen_kwargs["temperature"] = temperature

    total_len = 0
    for total_ids in model.stream_generate(**inputs, eos_token_id=eos_token_id, **gen_kwargs):
        total_ids = total_ids.tolist()[0]
        total_len = len(total_ids)
        if echo:
            output_ids = total_ids[:-1]
        else:
            output_ids = total_ids[input_echo_len:-1]

        response = tokenizer.decode(output_ids)
        if response and response[-1] != "<EFBFBD>":
            response, stop_found = apply_stopping_strings(response, ["<|observation|>"])

            yield {
                "text": response,
                "usage": {
                    "prompt_tokens": input_echo_len,
                    "completion_tokens": total_len - input_echo_len,
                    "total_tokens": total_len,
                },
                "finish_reason": "function_call" if stop_found else None,
            }

            if stop_found:
                break

    # Only last stream result contains finish_reason, we set finish_reason as stop
    ret = {
        "text": response,
        "usage": {
            "prompt_tokens": input_echo_len,
            "completion_tokens": total_len - input_echo_len,
            "total_tokens": total_len,
        },
        "finish_reason": "stop",
    }
    yield ret

    gc.collect()
    torch.cuda.empty_cache()


def process_chatglm_messages(messages, functions=None):
    _messages = messages
    messages = []

    if functions:
        messages.append(
            {
                "role": "system",
                "content": "Answer the following questions as best as you can. You have access to the following tools:",
                "tools": functions
            }
        )

    for m in _messages:
        role, content, func_call = m.role, m.content, m.function_call
        if role == "function":
            messages.append(
                {
                    "role": "observation",
                    "content": content
                }
            )

        elif role == "assistant" and func_call is not None:
            for response in content.split("<|assistant|>"):
                metadata, sub_content = response.split("\n", maxsplit=1)
                messages.append(
                    {
                        "role": role,
                        "metadata": metadata,
                        "content": sub_content.strip()
                    }
                )
        else:
            messages.append({"role": role, "content": content})
    return messages


def generate_chatglm3(model: PreTrainedModel, tokenizer: PreTrainedTokenizer, params: dict):
    for response in generate_stream_chatglm3(model, tokenizer, params):
        pass
    return response


def apply_stopping_strings(reply, stop_strings) -> Tuple[str, bool]:
    stop_found = False
    for string in stop_strings:
        idx = reply.find(string)
        if idx != -1:
            reply = reply[:idx]
            stop_found = True
            break

    if not stop_found:
        # If something like "\nYo" is generated just before "\nYou: is completed, trim it
        for string in stop_strings:
            for j in range(len(string) - 1, 0, -1):
                if reply[-j:] == string[:j]:
                    reply = reply[:-j]
                    break
            else:
                continue

            break

    return reply, stop_found