首页 > 编程语言 >ChatGLM3 源码解析(五)

ChatGLM3 源码解析(五)

时间:2024-03-13 13:45:43浏览次数:28  
标签:generation self ChatGLM3 ids length 源码 input 解析 config

PrefixEncoder

# 根据前缀 ID 获取前缀嵌入
# 前缀嵌入将连接到分头之后的 K 和 V 上
class PrefixEncoder(torch.nn.Module):
    """
    The torch.nn model to encode the prefix
    Input shape: (batch-size, prefix-length)
    Output shape: (batch-size, prefix-length, 2*layers*hidden)
    """

    def __init__(self, config: ChatGLMConfig):
        super().__init__()
        # 控制是否开启前缀投影,即用两层 MLP 处理前缀嵌入
        self.prefix_projection = config.prefix_projection
        if self.prefix_projection:
            # KVSize = NLayer * 2 * NGroup * HeadSize
            kv_size = config.num_layers * config.kv_channels * config.multi_query_group_num * 2
            # 将 ID 变为嵌入的嵌入层,[PreSeqLen, KVSize]
            self.embedding = torch.nn.Embedding(config.pre_seq_len, kv_size)
            # 处理嵌入的 MLP
            # 映射到 HidSize, 计算 tanh,在映射到 KVSize
            self.trans = torch.nn.Sequential(
                torch.nn.Linear(kv_size, config.hidden_size),
                torch.nn.Tanh(),
                torch.nn.Linear(config.hidden_size, kv_size)
            )
        else:
            # 将 ID 变为嵌入的嵌入层
            self.embedding = torch.nn.Embedding(config.pre_seq_len,
                                                config.num_layers * config.kv_channels * config.multi_query_group_num * 2)

    def forward(self, prefix: torch.Tensor):
        # 前缀 ID 尺寸为 [BatchSize, PreSeqLen]
        # 根据前缀 ID 获取嵌入,尺寸为 [BatchSize, PreSeqLen, KVSize]
        # 如果设定了需要投影,就用两层 MLP 处理嵌入
        if self.prefix_projection:
            prefix_tokens = self.embedding(prefix)
            past_key_values = self.trans(prefix_tokens)
        else:
            past_key_values = self.embedding(prefix)
        return past_key_values


ChatGLMPreTrainedModel

class ChatGLMPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and
    a simple interface for downloading and loading pretrained models.
    """

    is_parallelizable = False
    supports_gradient_checkpointing = True
    config_class = ChatGLMConfig
    base_model_prefix = "transformer"
    _no_split_modules = ["GLMBlock"]

    def _init_weights(self, module: nn.Module):
        """Initialize the weights."""
        return

    # 从输入单词 ID,KVCache生成默认的(上三角)掩码矩阵
    def get_masks(self, input_ids, past_key_values, padding_mask=None):
        # 单词 ID 尺寸为 [BatchSize, SeqLen]
        batch_size, seq_length = input_ids.shape
        # 掩码矩阵初始化为全 1,形状为 [BatchSize, SeqLen, SeqLen],每个输入序列一个
        full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
        # 保留其下三角元素,其余设为 9
        full_attention_mask.tril_()
        # CacheLen:KVCache 中序列长度
        # 如果没有提供则设为 0,如果提供了,从中获取长度
        past_length = 0
        if past_key_values:
            past_length = past_key_values[0][0].shape[0]
        # 如果提供了 KVCache,在每个掩码矩阵的上方填充 1,形状为 [BatchSize, SeqLen, CacheSeqLen]
        if past_length:
            full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
                                                        device=input_ids.device), full_attention_mask), dim=-1)

        # 如果提供了掩码数组([BatchSize, (Cache)SeqLen])
        # 将其变形为 [BatchSize, 1, (Cache)SeqLen]
        # 然后与掩码矩阵相乘
        # 将掩码数组为0的列设为0
        if padding_mask is not None:
            full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
        # 如果提供了掩码数组,并且没有提供 KVCache
        # 将其变形为 [BatchSize, SeqLen, 1]
        # 然后将掩码数组为 0 的行设为 1
        if not past_length and padding_mask is not None:
            full_attention_mask -= padding_mask.unsqueeze(-1) - 1
        # 小于 0.5 变成 true,大于 0.5 变成 false,相当于将其翻转,上三角不为 0
        full_attention_mask = (full_attention_mask < 0.5).bool()
        # 分头,变形为 [BatchSize, 1, SeqLen, SeqLen]
        full_attention_mask.unsqueeze_(1)
        return full_attention_mask

    # 从输入单词 ID 生成默认的(从零开始的)序列 ID
    def get_position_ids(self, input_ids, device):
        # 单词 ID 尺寸为 [BatchSize, SeqLen]
        batch_size, seq_length = input_ids.shape
        # 序列 ID 创建为 0~(SeqLen-1)的一维数组
        # 变形为 [1, SeqLen],之后重复第一维 BatchSize 次,得到 [BatchSize, SeqLen]
        position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
        return position_ids

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, GLMTransformer):
            module.gradient_checkpointing = value

ChatGLMForConditionalGeneration.stream_generate()

    @torch.inference_mode()
    def stream_generate(
            self,
            input_ids,
            generation_config: Optional[GenerationConfig] = None,
            logits_processor: Optional[LogitsProcessorList] = None,
            stopping_criteria: Optional[StoppingCriteriaList] = None,
            prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
            return_past_key_values=False,
            **kwargs,
    ):
        batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]

        if generation_config is None:
            generation_config = self.generation_config
        generation_config = copy.deepcopy(generation_config)
        model_kwargs = generation_config.update(**kwargs)
        model_kwargs["use_cache"] = generation_config.use_cache
        bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id

        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None

        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
        if has_default_max_length and generation_config.max_new_tokens is None:
            warnings.warn(
                f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
                "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
                " recommend using `max_new_tokens` to control the maximum length of the generation.",
                UserWarning,
            )
        elif generation_config.max_new_tokens is not None:
            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
            if not has_default_max_length:
                logger.warn(
                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
                    "Please refer to the documentation for more information. "
                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
                    UserWarning,
                )

        # 如果 SeqLen 大于等于配置里设定的 MaxSeqLen,发出警告
        if input_ids_seq_length >= generation_config.max_length:
            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
            logger.warning(
                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
                " increasing `max_new_tokens`."
            )

        # 如果没有提供 logits 处理器,初始化为空列表
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        # 没有提供停止标准,初始化为空列表
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()

        # 根据生成配置等对象获取 logits 处理器
        logits_processor = self._get_logits_processor(
            generation_config=generation_config,
            input_ids_seq_length=input_ids_seq_length,
            encoder_input_ids=input_ids,
            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
            logits_processor=logits_processor,
        )
        # 根据生成配置等对象获取停止标准
        stopping_criteria = self._get_stopping_criteria(
            generation_config=generation_config, stopping_criteria=stopping_criteria
        )
        # 根据生成配置获取 logits 包装器
        logits_warper = self._get_logits_warper(generation_config)

        # 未完成标志,表示每个序列是否生成完毕的数组
        # 初始化为 [BatchSize] 尺寸的全 1 数组
        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
        scores = None
        while True:
            # 根据传入参数组装成字典,请见该方法定义
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
            # 将单词 ID 传入模型,得到(所有前缀)下一个单词的 logits
            # [BatchSize, SeqLen, VocabSize]
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=False,
                output_hidden_states=False,
            )
            # 截取 SeqLen 维度的最后一维,得到整句话下一个单词的 logits
            # [BatchSize, VocabSize]
            next_token_logits = outputs.logits[:, -1, :]

            # 传入 logits 处理器和包装器,修正 logits
            next_token_scores = logits_processor(input_ids, next_token_logits)
            next_token_scores = logits_warper(input_ids, next_token_scores)

            # 计算 softmax 得到概率值
            probs = nn.functional.softmax(next_token_scores, dim=-1)
            # 如果设定了需要采样,对其进行多项式采样,样本容量为 1
            # 否则直接取最大的
            # 得到下个单词 ID,尺寸为 [BatchSize]
            if generation_config.do_sample:
                next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
            else:
                next_tokens = torch.argmax(probs, dim=-1)
            # 下个单词 ID 变形为 [BatchSize, 1],然后和输入单词 ID 拼接
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
            # 根据当前输出更新KVCache、注意力掩码和位置ID
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
            # `next_tokens` 变形为 [1, BatchSize],再将第一维重复 NEOS 次,[NEOS, BatchSize]
            # `eos_token_id_tensor` 变形为 [NEOS, 1],将广播第二维变成 [NEOS, BatchSize]
            # 之后二者逐元素比较是否不相等,形成一个比较结果,尺寸为 [NEOS, BatchSize]
            # 之后按照 BatchSize 维度计算乘积,得到未完成标志,[BatchSize]
            # 如果某个序列等于终止符集合里面的任意一个,那么比较结果就会出现一个 0,未完成标志将会是 0。
            unfinished_sequences = unfinished_sequences.mul(
                next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
            )
            # 如果指定了返回 KVCache
            # 产生输入ID和已生成的输出ID
            # 和 KVCache
            # 否则只产生第一个
            if return_past_key_values:
                yield input_ids, outputs.past_key_values
            else:
                yield input_ids
            # 如果未完成标志全为零(表示序列都已生成完毕),或者达到了停止标准,就停止生成
            if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
                break

标签:generation,self,ChatGLM3,ids,length,源码,input,解析,config
From: https://www.cnblogs.com/apachecn/p/18070429

相关文章