2023年的深度学习入门指南(26) - 在自己电脑上运行通义千问7b模型

2023年的深度学习入门指南(26) - 在自己电脑上运行通义千问7b模型

通过量化，通义千问4位量化的模型大小为5.86G，可以在3060等小于16G的家用GPU上也可以运行起来。

通义千问7b的量化运行

通义千问7b提供了4位量化好的Qwen/Qwen-7B-Chat-Int4模型，我们直接调用就好。

首先安装依赖包：

pip install transformers==4.32.0
pip install accelerate
pip install tiktoken
pip install einops
pip install transformers_stream_generator==0.0.4
pip install scipy
pip install auto-gptq optimum

如果你是Linux环境的话，可以安装下Flash-Attention来加速：

git clone -b v1.0.8 https://github.com/Dao-AILab/flash-attention
cd flash-attention && pip install .

Windows下暂时还用不了，这个不是必选步骤。

下面我们就可以来写代码调用通义千问7b了：

from transformers import AutoTokenizer, AutoModelForCausalLM

# Note: The default behavior now has injection attack prevention off.
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-7B-Chat-Int4", trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen-7B-Chat-Int4",
    device_map="auto",
    trust_remote_code=True
).eval()
response, history = model.chat(tokenizer, "生成用C++将字符串倒序的代码", history=None)
print(response)

生成结果如下：

以下是C++中将字符串逆序的示例代码：


#include <iostream>
#include <string>

int main() {
    
    
    std::string str = "Hello, World!";
    std::string reversedStr = str;
    std::reverse(reversedStr.begin(), reversedStr.end());
    std::cout << reversedStr << std::endl;
    return 0;
}


首先，我们定义了一个包含字符串的变量 `str`。然后，我们定义了一个空字符串变量 `reversedStr`，用于存储逆序后的字符串。

接下来，我们使用 `std::reverse()` 函数将 `str` 中的字符逆序。该函数需要一个迭代器范围作为参数，表示要逆序的字符序列。在这里，我们使用 `str.begin()` 和 `str.end()` 获取字符串的起始和结束迭代器，然后将它们传递给 `std::reverse()` 函数。

最后，我们输出逆序后的字符串。

我是在3060 GPU上运行成功的。

下面我们继续讲解通义千问7B的源代码。

通义千问7b的全连接网络

除了使用了silu激活函数之外，其他就是基本的全连接网络了。

class QWenMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.w1 = nn.Linear(
            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
        )
        self.w2 = nn.Linear(
            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
        )
        ff_dim_in = config.intermediate_size // 2
        self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)

    def forward(self, hidden_states):
        a1 = self.w1(hidden_states)
        a2 = self.w2(hidden_states)
        intermediate_parallel = a1 * F.silu(a2)
        output = self.c_proj(intermediate_parallel)
        return output

SiLU 函数是一种神经网络中的激活函数，全称是 Sigmoid Linear Unit, 也被称为 Swish 函数。它由 Google Brain 在 2017 年提出，是一种非线性激活函数，能够有效地对神经网络的输入进行非线性变换。

SiLU 函数的定义如下：

f(x) = x * sigmoid(x)

其中，sigmoid 函数是 Sigmoid 函数，定义如下：

sigmoid(x) = 1 / (1 + exp(-x))

SiLU 函数的特点如下：

• 正数区域内，SiLU 函数的输出与 ReLU 函数的输出相同。
• 在负数区域内，SiLU 函数的输出与 sigmoid 函数的输出相同。
• SiLU 函数在整个定义域内都是可微的，这使得在反向传播过程中的梯度计算更加稳定。
• SiLU函数不是单调递增的，而是在x≈−1.28时达到全局最小值−0.28，这可以起到一个隐式正则化的作用，抑制过大的权重

Transformer块

下面我们将RMSNorm，QWenAttention和QWenMLP三者搭建成QWenBlock，就类似于LLaMA中的TransformerBlock：

class QWenBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        hidden_size = config.hidden_size
        self.bf16 = config.bf16

        self.ln_1 = RMSNorm(
            hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.attn = QWenAttention(config)
        self.ln_2 = RMSNorm(
            hidden_size,
            eps=config.layer_norm_epsilon,
        )

        self.mlp = QWenMLP(config)

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        rotary_pos_emb: Optional[List[torch.Tensor]] = None,
        registered_causal_mask: Optional[torch.Tensor] = None,
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ):
        layernorm_output = self.ln_1(hidden_states)

        attn_outputs = self.attn(
            layernorm_output,
            rotary_pos_emb,
            registered_causal_mask=registered_causal_mask,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attn_outputs[0]

        outputs = attn_outputs[1:]

        residual = hidden_states
        layernorm_input = attn_output + residual

        layernorm_output = self.ln_2(layernorm_input)

        residual = layernorm_input
        mlp_output = self.mlp(layernorm_output)
        hidden_states = residual + mlp_output

        if use_cache:
            outputs = (hidden_states,) + outputs
        else:
            outputs = (hidden_states,) + outputs[1:]

        return outputs

这一模块主要就是将一些参数传递给上节我们介绍过的QWenAttention:

• hidden_states：一个可选的元组，包含了上一层的输出张量，形状为(batch_size, sequence_length, hidden_size)。
• rotary_pos_emb：一个可选的列表，包含了旋转位置编码张量，形状为(batch_size, sequence_length, hidden_size)。
• registered_causal_mask：一个可选的张量，用于注册因果掩码，防止模型看到未来的信息。形状为(batch_size, sequence_length, sequence_length)。
• layer_past：一个可选的元组，包含了上一层的注意力键值对张量，用于实现缓存机制，加速生成过程。形状为(2, batch_size, num_heads, sequence_length, head_dim)。
• attention_mask：一个可选的浮点张量，用于对输入序列进行掩码，忽略无效的位置或填充部分。形状为(batch_size, sequence_length)或(batch_size, 1, 1, sequence_length)。
• head_mask：一个可选的浮点张量，用于对注意力头进行掩码，随机删除一些头以增加模型的鲁棒性。形状为(num_heads,)或(1, 1, num_heads, 1)。
• encoder_hidden_states：一个可选的张量，用于实现编码器-解码器结构时，传递编码器的输出给解码器。形状为(batch_size, encoder_sequence_length, hidden_size)。
• encoder_attention_mask：一个可选的浮点张量，用于实现编码器-解码器结构时，对编码器输出进行掩码。形状为(batch_size, encoder_sequence_length)或(batch_size, 1, 1, encoder_sequence_length)。
• use_cache：一个可选的布尔值，用于指示是否使用缓存机制。
• output_attentions：一个可选的布尔值，用于指示是否输出注意力权重张量。

RMSNorm

RMSNorm我们已经讲过多次的，这里就不多介绍了：

class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        if rms_norm is not None and x.is_cuda:
            return rms_norm(x, self.weight, self.eps)
        else:
            output = self._norm(x.float()).type_as(x)
            return output * self.weight

位置编码

还记得讲百川模型代码时我们遇到的einsum吗？在千问的代码里我们会再次遇到这样的爱因斯坦风格，这次我们用到的是一个库einops。

在einops的加持下，我们可以将维度变换的操作变得更有可读性：

            from einops import rearrange

            emb = rearrange(emb, "n d -> 1 n 1 d")

rearrange函数可以根据字符串表达式来重新排列张量维度。

这里的"n d -> 1 n 1 d"表示:

• 从(n, d)形状
• 重新排列为(1, n, 1, d)形状
  也就是在emb张量的维度1(n个向量)前面增加两维,变成1和1。

其余的还是使用cos和sin函数作cache:

class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim, base=10000):
        super().__init__()
        self.dim = dim
        self.base = base
        self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        if importlib.util.find_spec("einops") is None:
            raise RuntimeError("einops is required for Rotary Embedding")

        self._rotary_pos_emb_cache = None
        self._seq_len_cached = 0
        self._ntk_alpha_cached = 1.0

    def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
        seqlen = max_seq_len + offset
        if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
            base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
            self.inv_freq = 1.0 / (
                base
                ** (
                    torch.arange(0, self.dim, 2, device=self.inv_freq.device).float()
                    / self.dim
                )
            )
            self._seq_len_cached = max(2 * seqlen, 16)
            self._ntk_alpha_cached = ntk_alpha
            seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
            freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
            
            emb = torch.cat((freqs, freqs), dim=-1)
            from einops import rearrange

            emb = rearrange(emb, "n d -> 1 n 1 d")

            cos, sin = emb.cos(), emb.sin()
            self._rotary_pos_emb_cache = [cos, sin]

    def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
        self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
        cos, sin = self._rotary_pos_emb_cache
        return [cos[:, offset : offset + max_seq_len], sin[:, offset : offset + max_seq_len]]

千问7B的旋转函数也是用einops.rearrange来实现的:

def _rotate_half(x):
    from einops import rearrange

    x = rearrange(x, "... (j d) -> ... j d", j=2)
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1)

最后是apply_rotary_pos_emb函数,作用是将旋转位置编码应用到输入张量t上。

def apply_rotary_pos_emb(t, freqs):
    cos, sin = freqs
    if apply_rotary_emb_func is not None and t.is_cuda:
        t_ = t.float()
        cos = cos.squeeze(0).squeeze(1)[:, : cos.shape[-1] // 2]
        sin = sin.squeeze(0).squeeze(1)[:, : sin.shape[-1] // 2]
        output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
        return output
    else:
        rot_dim = freqs[0].shape[-1]
        cos, sin = freqs
        t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
        t_ = t_.float()
        t_pass_ = t_pass_.float()
        t_ = (t_ * cos) + (_rotate_half(t_) * sin)
        return torch.cat((t_, t_pass_), dim=-1).type_as(t)

apply_rotary_pos_emb的主要步骤:

• 从freqs中分离出cos和sin编码。
• 如果CUDA环境且有apply_rotary_emb_func实现,直接调用该函数进行优化的旋转编码。
• 否则,手动实现旋转编码:
• 将t切分为要编码部分t_和不编码部分t_pass_。
• 计算旋转编码后的t_。
• 将编码后的t_和未编码的t_pass_拼接。
• 返回拼接后的结果。

这样,当有优化实现时直接调用,否则用Python实现旋转位置编码。

旋转位置编码的作用是让模型表征更具局部性,使自注意力更聚焦在关键区域。这通常能提升长序列建模的性能。

通义千问的Transformer模型

class QWenModel(QWenPreTrainedModel):
    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]

    def __init__(self, config):
        super().__init__(config)
        self.vocab_size = config.vocab_size
        self.num_hidden_layers = config.num_hidden_layers
        self.embed_dim = config.hidden_size

        self.gradient_checkpointing = False
        self.use_dynamic_ntk = config.use_dynamic_ntk
        self.seq_length = config.seq_length

        self.wte = nn.Embedding(self.vocab_size, self.embed_dim)

        self.drop = nn.Dropout(config.emb_dropout_prob)

        if config.rotary_pct == 1.0:
            self.rotary_ndims = None
        else:
            assert config.rotary_pct < 1
            self.rotary_ndims = int(
                config.kv_channels * config.rotary_pct
            )
        dim = (
            self.rotary_ndims
            if self.rotary_ndims is not None
            else config.kv_channels
        )
        self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)

        self.use_flash_attn = config.use_flash_attn
        self.is_fp32 = not (config.bf16 or config.fp16)
        if (
            self.use_flash_attn
            and flash_attn_unpadded_func is not None
            and not self.is_fp32
        ):
            self.registered_causal_mask = None
        else:
            max_positions = config.max_position_embeddings
            self.register_buffer(
                "registered_causal_mask",
                torch.tril(
                    torch.ones((max_positions, max_positions), dtype=torch.bool)
                ).view(1, 1, max_positions, max_positions),
                persistent=False,
            )

        self.h = nn.ModuleList(
            [
                QWenBlock(
                    config
                )
                for i in range(config.num_hidden_layers)
            ]
        )
        self.ln_f = RMSNorm(
            self.embed_dim,
            eps=config.layer_norm_epsilon,
        )

        self.post_init()

初始化的部分还是将之前介绍过的各模块组合在一起。

下面是虽然大但是主要是例行公事和错误判断的forward:

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, input_shape[-1])
        if position_ids is not None:
            position_ids = position_ids.view(-1, input_shape[-1])

        if past_key_values is None:
            past_length = 0
            past_key_values = tuple([None] * len(self.h))
        else:
            past_length = past_key_values[0][0].size(-2)

        if position_ids is None:
            position_ids = torch.arange(
                past_length,
                input_shape[-1] + past_length,
                dtype=torch.long,
                device=device,
            )
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        if attention_mask is not None:
            if batch_size <= 0:
                raise ValueError("batch_size has to be defined and > 0")
            attention_mask = attention_mask.view(batch_size, -1)
            attention_mask = attention_mask[:, None, None, :]
            attention_mask = attention_mask.to(dtype=self.dtype)
            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min

        encoder_attention_mask = None
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)
        hidden_states = inputs_embeds

        kv_seq_len = hidden_states.size()[1]
        if past_key_values[0] is not None:
            # past key values[0][0] shape: bs * seq_len * head_num * dim
            kv_seq_len += past_key_values[0][0].shape[1]
        if (
            self.use_dynamic_ntk
            and kv_seq_len == hidden_states.size()[1]
            and not self.training
        ):
            context_value = math.log(kv_seq_len / self.seq_length, 2) + 1
            ntk_alpha = 2 ** math.ceil(context_value) - 1
            ntk_alpha = max(ntk_alpha, 1)
        else:
            ntk_alpha = self.rotary_emb._ntk_alpha_cached

        rotary_pos_emb = self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha)
        for idx in range(len(rotary_pos_emb)):
            rotary_pos_emb[idx] = rotary_pos_emb[idx].to(hidden_states.device)

        hidden_states = self.drop(hidden_states)
        output_shape = input_shape + (hidden_states.size(-1),)

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, use_cache, output_attentions)

                    return custom_forward

                outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    rotary_pos_emb,
                    self.registered_causal_mask,
                    None,
                    attention_mask,
                    head_mask[i],
                    encoder_hidden_states,
                    encoder_attention_mask,
                )
            else:
                outputs = block(
                    hidden_states,
                    layer_past=layer_past,
                    rotary_pos_emb=rotary_pos_emb,
                    registered_causal_mask=self.registered_causal_mask,
                    attention_mask=attention_mask,
                    head_mask=head_mask[i],
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)

        hidden_states = self.ln_f(hidden_states)
        hidden_states = hidden_states.view(output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, presents, all_hidden_states] if v is not None
            )

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )

这实现了一个标准的Transformer编码器结构,有输入处理、Encoding块循环、输出后处理三个主要部分。使用了层规范化、多头自注意力、残差连接等机制。还支持caching、checkpoints、mask等功能。

预训练模型

下面再说一下QWenModel的基类，用于设置并行训练和保存点等信息的，继承自PreTrainedModel的类：

class QWenPreTrainedModel(PreTrainedModel):
    config_class = QWenConfig
    base_model_prefix = "transformer"
    is_parallelizable = False
    supports_gradient_checkpointing = True
    _no_split_modules = ["QWenBlock"]

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, RMSNorm):
            module.weight.data.fill_(1.0)

        for name, p in module.named_parameters():
            if name == "c_proj.weight":
                p.data.normal_(
                    mean=0.0,
                    std=(
                        self.config.initializer_range
                        / math.sqrt(2 * self.config.num_hidden_layers)
                    ),
                )

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

语言模型封装

上面的QWenModel返回的BaseModelOutputWithPast，如果要做成语言模型的话，还要封装成CausalLMOutputWithPast。

class QWenLMHeadModel(QWenPreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
    _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]

    def __init__(self, config):
        super().__init__(config)
        assert (
            config.bf16 + config.fp16 + config.fp32 <= 1
        ), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"

        autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0

        if autoset_precision:
            if SUPPORT_BF16:
                logger.warn(
                    "The model is automatically converting to bf16 for faster inference. "
                    "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
                )
                config.bf16 = True
            elif SUPPORT_FP16:
                logger.warn(
                    "The model is automatically converting to fp16 for faster inference. "
                    "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
                )
                config.fp16 = True
            else:
                config.fp32 = True

        if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16:
            logger.warn("Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".")
        if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16:
            logger.warn("Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster")
        if config.fp32:
            if SUPPORT_BF16:
                logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
            elif SUPPORT_FP16:
                logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
        
        if config.use_flash_attn == "auto":
            if config.bf16 or config.fp16:
                logger.warn("Try importing flash-attention for faster inference...")
                config.use_flash_attn = True
            else:
                config.use_flash_attn = False
        if config.use_flash_attn and config.fp32:
            logger.warn("Flash attention will be disabled because it does NOT support fp32.")

        if config.use_flash_attn:
            _import_flash_attn()

        self.transformer = QWenModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        if config.bf16:
            self.transformer.bfloat16()
            self.lm_head.bfloat16()
        if config.fp16:
            self.transformer.half()
            self.lm_head.half()
        self.post_init()

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
    ):
        token_type_ids = kwargs.get("token_type_ids", None)
        if past_key_values:
            input_ids = input_ids[:, -1].unsqueeze(-1)
            if token_type_ids is not None:
                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)

        attention_mask = kwargs.get("attention_mask", None)
        position_ids = kwargs.get("position_ids", None)

        if attention_mask is not None and position_ids is None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -1].unsqueeze(-1)
        else:
            position_ids = None

        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {
    
    "inputs_embeds": inputs_embeds}
        else:
            model_inputs = {
    
    "input_ids": input_ids}

        model_inputs.update(
            {
    
    
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "position_ids": position_ids,
                "attention_mask": attention_mask,
                "token_type_ids": token_type_ids,
            }
        )
        return model_inputs

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            labels = labels.to(lm_logits.device)
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
            )

        if not return_dict:
            output = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )

在forward之外，语言模型还需要封装一个生成函数。主要也是做一些配置，然后调用父类的生成函数：

    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
        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,
        synced_gpus: Optional[bool] = None,
        assistant_model: Optional["PreTrainedModel"] = None,
        streamer: Optional["BaseStreamer"] = None,
        **kwargs,
    ) -> Union[GenerateOutput, torch.LongTensor]:
        generation_config = generation_config if generation_config is not None else self.generation_config

        # Process stop_words_ids.
        stop_words_ids = kwargs.pop("stop_words_ids", None)
        if stop_words_ids is None and generation_config is not None:
            stop_words_ids = getattr(generation_config, "stop_words_ids", None)
        if stop_words_ids is None:
            stop_words_ids = getattr(generation_config, "stop_words_ids", None)

        if stop_words_ids is not None:
            stop_words_logits_processor = StopWordsLogitsProcessor(
                stop_words_ids=stop_words_ids,
                eos_token_id=generation_config.eos_token_id,
            )
            if logits_processor is None:
                logits_processor = LogitsProcessorList([stop_words_logits_processor])
            else:
                logits_processor.append(stop_words_logits_processor)

        return super().generate(
            inputs,
            generation_config=generation_config,
            logits_processor=logits_processor,
            stopping_criteria=stopping_criteria,
            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
            synced_gpus=synced_gpus,
            assistant_model=assistant_model,
            streamer=streamer,
            **kwargs,
        )

聊天功能封装

    def chat(
        self,
        tokenizer: PreTrainedTokenizer,
        query: str,
        history: Optional[HistoryType],
        system: str = "You are a helpful assistant.",
        append_history: bool = True,
        stream: Optional[bool] = _SENTINEL,
        stop_words_ids: Optional[List[List[int]]] = None,
        generation_config: Optional[GenerationConfig] = None,
        **kwargs,
    ) -> Tuple[str, HistoryType]:
        generation_config = generation_config if generation_config is not None else self.generation_config

        assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
        if history is None:
            history = []
        if stop_words_ids is None:
            stop_words_ids = []

        max_window_size = kwargs.get('max_window_size', None)
        if max_window_size is None:
            max_window_size = generation_config.max_window_size
        raw_text, context_tokens = make_context(
            tokenizer,
            query,
            history=history,
            system=system,
            max_window_size=max_window_size,
            chat_format=generation_config.chat_format,
        )

        stop_words_ids.extend(get_stop_words_ids(
            generation_config.chat_format, tokenizer
        ))
        input_ids = torch.tensor([context_tokens]).to(self.device)
        outputs = self.generate(
                    input_ids,
                    stop_words_ids=stop_words_ids,
                    return_dict_in_generate=False,
                    generation_config=generation_config,
                    **kwargs,
                )

        response = decode_tokens(
            outputs[0],
            tokenizer,
            raw_text_len=len(raw_text),
            context_length=len(context_tokens),
            chat_format=generation_config.chat_format,
            verbose=False,
            errors='replace'
        )

        if append_history:
            history.append((query, response))

        return response, history

其主要流程如下：

流式聊天封装

最后是封装成可以流式获取的函数。

其主要流程为：

• 和chat方法类似,先做输入query的处理,组装context。
• 计算停止词stop_words_ids。
• 将停止词集合封装成StopWordsLogitsProcessor。
• 将context转成input_ids作为模型输入。
• 关键在这里,调用generate_stream方法进行流式生成。它会逐个token地生成序列,并用yield返回每个结果。
• 在一个while循环中收集生成的token,并用decode方法转成文本。
• 通过yield关键字返回每个解码的结果。
• 最终形成一个生成器,可以不断获取模型生成的内容。

    def chat_stream(
            self,
            tokenizer: PreTrainedTokenizer,
            query: str,
            history: Optional[HistoryType],
            system: str = "You are a helpful assistant.",
            stop_words_ids: Optional[List[List[int]]] = None,
            logits_processor: Optional[LogitsProcessorList] = None,
            generation_config: Optional[GenerationConfig] = None,
            **kwargs,
    ) -> Generator[str, Any, None]:
        generation_config = generation_config if generation_config is not None else self.generation_config
        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
        if history is None:
            history = []
        if stop_words_ids is None:
            stop_words_ids = []

        max_window_size = kwargs.get('max_window_size', None)
        if max_window_size is None:
            max_window_size = generation_config.max_window_size
        raw_text, context_tokens = make_context(
            tokenizer,
            query,
            history=history,
            system=system,
            max_window_size=max_window_size,
            chat_format=generation_config.chat_format,
        )

        stop_words_ids.extend(get_stop_words_ids(
            generation_config.chat_format, tokenizer
        ))
        if stop_words_ids is not None:
            stop_words_logits_processor = StopWordsLogitsProcessor(
                stop_words_ids=stop_words_ids,
                eos_token_id=generation_config.eos_token_id,
            )
            if logits_processor is None:
                logits_processor = LogitsProcessorList([stop_words_logits_processor])
            else:
                logits_processor.append(stop_words_logits_processor)
        input_ids = torch.tensor([context_tokens]).to(self.device)

        from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
        self.__class__.generate_stream = NewGenerationMixin.generate
        self.__class__.sample_stream = NewGenerationMixin.sample_stream
        stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)

        def stream_generator():
            outputs = []
            for token in self.generate_stream(
                    input_ids,
                    return_dict_in_generate=False,
                    generation_config=stream_config,
                    logits_processor=logits_processor,
                    seed=-1,
                    **kwargs):
                outputs.append(token.item())
                yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore')

        return stream_generator()

小结

这节我们终于介绍完了千问7b的模型的代码。凡是讲源码的肯定会遇到大量细节，这些细节也未必是值得花太多精力去抠的，但是原汁原味的代码还是能更精确地表达功能的真实含义。
后面我们还会将模型实现抽象一下，做更系统化的讲解便于初学者理解。对于从业的同学，因为你们面对的就是这些细节，所以先熟悉起来吧。