GPT(1)----------从零实现 GPT 模型以生成文本

小女孩真可爱2026-08-25 10:24

本文用于记录和学习GPT搭建，以及一个训练优化的技巧，对于基础知识需自行去补充。

1. 概述

本文档描述一个完整的 GPT (Generative Pre-trained Transformer) 语言模型实现。该实现对应 **GPT-2 Small(124M 参数)** 的标准结构,涵盖从 token 输入到生成文本的完整流程。

模型采用 **仅解码器(Decoder-only)** 的 Transformer 架构,通过因果自注意力(Causal Self-Attention)实现自回归文本生成。

2. 整体架构

模型的前向计算数据流如下:

输入 token 索引 (batch, seq_len)

Token Embedding + Positional Embedding

▼

Dropout

▼

N × TransformerBlock (堆叠 n_layers 层)

▼

Final LayerNorm

▼

输出线性层 (emb_dim → vocab_size)

▼

logits (batch, seq_len, vocab_size)

3. 核心组件

3.1 嵌入层 (Embedding)

| 组件 | 作用 |

| `tok_emb` | 将每个 token id 映射为 `emb_dim` 维向量 |

| `pos_emb` | 为每个位置提供位置向量,使模型感知词序 |

| `drop_emb` | 对嵌入结果做 Dropout 正则化 |

Token 嵌入与位置嵌入 **逐元素相加** 后进入后续网络

3.2 Transformer 块 (TransformerBlock)

模型核心,重复堆叠 `n_layers` 次(`01gpt.py:115`)。每个块包含两个子层,均采用 **Pre-LayerNorm + 残差连接** 结构:

```

x → LayerNorm → 多头注意力 → Dropout → +x (残差)

→ LayerNorm → 前馈网络 → Dropout → +x (残差)

```

  • **Pre-LN**:先归一化再进入子层,是 GPT-2 采用的方式,有助于训练稳定性。

  • **残差连接**:缓解深层网络的梯度消失问题。

3.3 多头注意力 (MultiHeadAttention)

**关键机制:**

  1. **Q/K/V 投影**:三个线性层分别生成 Query、Key、Value。

  2. **多头拆分**:将 `d_out` 拆成 `num_heads` 个头并行计算,每个头维度为 `head_dim = d_out / num_heads`。

  3. **因果掩码**:使用上三角掩码(`torch.triu`,`01gpt.py:30`)遮挡未来 token,保证每个位置只能关注自身及之前的 token。

  4. **缩放点积注意力**:注意力分数除以 `√head_dim` 进行缩放(`01gpt.py:61`),防止 softmax 梯度过小。

  5. **输出投影**:合并所有头后经过 `out_proj` 线性层。

计算公式:

```

Attention(Q, K, V) = softmax( (Q·Kᵀ) / √d_k + mask ) · V

```

3.4 前馈网络 (FeedForward)

代码位置:`01gpt.py:102`。

结构为「扩展---激活---压缩」:

```

Linear(emb_dim → 4·emb_dim) → GELU → Linear(4·emb_dim → emb_dim)

```

中间层扩展到 4 倍维度,提升模型表达能力。

3.5 层归一化 (LayerNorm)

代码位置:`01gpt.py:77`。

对最后一维做归一化,含可学习的缩放参数 `scale` 和平移参数 `shift`:

```

LN(x) = scale · (x - μ) / √(σ² + ε) + shift

```

其中 `ε = 1e-5`,方差使用有偏估计(`unbiased=False`)。

4. 模型配置 (GPT-2 124M)

| 参数 | 值 | 含义 |

|------|-----|------|

| `vocab_size` | 50257 | 词表大小(GPT-2 BPE) |

| `context_length` | 1024 | 最大上下文长度 |

| `emb_dim` | 768 | 嵌入 / 隐藏维度 |

| `n_heads` | 12 | 注意力头数 |

| `n_layers` | 12 | Transformer 层数 |

| `drop_rate` | 0.1 | Dropout 比率 |

| `qkv_bias` | False | Q/K/V 是否使用偏置 |

代码实现方式如下：（由于该模型没有训练，所以输出是随机无意义的的数据）

python 复制代码 
import time
import tiktoken
import torch
import torch.nn as nn


#####################################
# Chapter 3
#####################################
class MultiHeadAttention(nn.Module):
    def __init__(self, d_in, d_out, context_length, dropout, num_heads, qkv_bias=False):
        super().__init__()
        assert d_out % num_heads == 0, "d_out must be divisible by num_heads"

        self.d_out = d_out
        self.num_heads = num_heads
        self.head_dim = d_out // num_heads  # Reduce the projection dim to match desired output dim

        self.W_query = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.W_key = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.W_value = nn.Linear(d_in, d_out, bias=qkv_bias)
        self.out_proj = nn.Linear(d_out, d_out)  # Linear layer to combine head outputs
        self.dropout = nn.Dropout(dropout)
        self.register_buffer(
            "mask",
            torch.triu(torch.ones(context_length, context_length), diagonal=1),
            persistent=False
        )

    def forward(self, x):
        b, num_tokens, d_in = x.shape

        keys = self.W_key(x)  # Shape: (b, num_tokens, d_out)
        values = self.W_value(x)
        queries = self.W_query(x)

        # We implicitly split the matrix by adding a `num_heads` dimension
        # Unroll last dim: (b, num_tokens, d_out) -> (b, num_tokens, num_heads, head_dim)
        keys = keys.view(b, num_tokens, self.num_heads, self.head_dim)
        values = values.view(b, num_tokens, self.num_heads, self.head_dim)
        queries = queries.view(b, num_tokens, self.num_heads, self.head_dim)

        # Transpose: (b, num_tokens, num_heads, head_dim) -> (b, num_heads, num_tokens, head_dim)
        keys = keys.transpose(1, 2)
        queries = queries.transpose(1, 2)
        values = values.transpose(1, 2)

        # Compute scaled dot-product attention (aka self-attention) with a causal mask
        attn_scores = queries @ keys.transpose(2, 3)  # Dot product for each head

        # Original mask truncated to the number of tokens and converted to boolean
        mask_bool = self.mask.bool()[:num_tokens, :num_tokens]

        # Use the mask to fill attention scores
        attn_scores.masked_fill_(mask_bool, -torch.inf)

        attn_weights = torch.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)
        attn_weights = self.dropout(attn_weights)

        # Shape: (b, num_tokens, num_heads, head_dim)
        context_vec = (attn_weights @ values).transpose(1, 2)

        # Combine heads, where self.d_out = self.num_heads * self.head_dim
        context_vec = context_vec.contiguous().view(b, num_tokens, self.d_out)
        context_vec = self.out_proj(context_vec)  # optional projection

        return context_vec


#####################################
# Chapter 4
#####################################
class LayerNorm(nn.Module):
    def __init__(self, emb_dim):
        super().__init__()
        self.eps = 1e-5
        self.scale = nn.Parameter(torch.ones(emb_dim))
        self.shift = nn.Parameter(torch.zeros(emb_dim))

    def forward(self, x):
        mean = x.mean(dim=-1, keepdim=True)
        var = x.var(dim=-1, keepdim=True, unbiased=False)
        norm_x = (x - mean) / torch.sqrt(var + self.eps)
        return self.scale * norm_x + self.shift


class GELU(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return 0.5 * x * (1 + torch.tanh(
            torch.sqrt(torch.tensor(2.0 / torch.pi)) *
            (x + 0.044715 * torch.pow(x, 3))
        ))


class FeedForward(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.layers = nn.Sequential(
            nn.Linear(cfg["emb_dim"], 4 * cfg["emb_dim"]),
            GELU(),
            nn.Linear(4 * cfg["emb_dim"], cfg["emb_dim"]),
        )

    def forward(self, x):
        return self.layers(x)


class TransformerBlock(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.att = MultiHeadAttention(
            d_in=cfg["emb_dim"],
            d_out=cfg["emb_dim"],
            context_length=cfg["context_length"],
            num_heads=cfg["n_heads"],
            dropout=cfg["drop_rate"],
            qkv_bias=cfg["qkv_bias"])
        self.ff = FeedForward(cfg)
        self.norm1 = LayerNorm(cfg["emb_dim"])
        self.norm2 = LayerNorm(cfg["emb_dim"])
        self.drop_shortcut = nn.Dropout(cfg["drop_rate"])

    def forward(self, x):
        # Shortcut connection for attention block
        shortcut = x
        x = self.norm1(x)
        x = self.att(x)   # Shape [batch_size, num_tokens, emb_size]
        x = self.drop_shortcut(x)
        x = x + shortcut  # Add the original input back

        # Shortcut connection for feed-forward block
        shortcut = x
        x = self.norm2(x)
        x = self.ff(x)
        x = self.drop_shortcut(x)
        x = x + shortcut  # Add the original input back

        return x


class GPTModel(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.tok_emb = nn.Embedding(cfg["vocab_size"], cfg["emb_dim"])
        self.pos_emb = nn.Embedding(cfg["context_length"], cfg["emb_dim"])
        self.drop_emb = nn.Dropout(cfg["drop_rate"])

        self.trf_blocks = nn.Sequential(
            *[TransformerBlock(cfg) for _ in range(cfg["n_layers"])])

        self.final_norm = LayerNorm(cfg["emb_dim"])
        self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False)

    def forward(self, in_idx):
        batch_size, seq_len = in_idx.shape
        tok_embeds = self.tok_emb(in_idx)
        pos_embeds = self.pos_emb(torch.arange(seq_len, device=in_idx.device))
        x = tok_embeds + pos_embeds  # Shape [batch_size, num_tokens, emb_size]
        x = self.drop_emb(x)
        x = self.trf_blocks(x)
        x = self.final_norm(x)
        logits = self.out_head(x)
        return logits


def generate_text_simple(model, idx, max_new_tokens, context_size):
    model.eval()
    # idx is (B, T) array of indices in the current context
    for _ in range(max_new_tokens):

        # Crop current context if it exceeds the supported context size
        # E.g., if LLM supports only 5 tokens, and the context size is 10
        # then only the last 5 tokens are used as context
        idx_cond = idx[:, -context_size:]

        # Get the predictions
        with torch.no_grad():
            logits = model(idx_cond)

        # Focus only on the last time step
        # (batch, n_token, vocab_size) becomes (batch, vocab_size)
        logits = logits[:, -1, :]

        # Get the idx of the vocab entry with the highest logits value
        idx_next = torch.argmax(logits, dim=-1, keepdim=True)  # (batch, 1)

        # Append sampled index to the running sequence
        idx = torch.cat((idx, idx_next), dim=1)  # (batch, n_tokens+1)

    return idx


def main():
    GPT_CONFIG_124M = {
        "vocab_size": 50257,     # Vocabulary size
        "context_length": 1024,  # Context length
        "emb_dim": 768,          # Embedding dimension
        "n_heads": 12,           # Number of attention heads
        "n_layers": 12,          # Number of layers
        "drop_rate": 0.1,        # Dropout rate
        "qkv_bias": False        # Query-Key-Value bias
    }

    torch.manual_seed(123)
    model = GPTModel(GPT_CONFIG_124M)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    model.eval()  # disable dropout

    start_context = "Hello, I am"

    tokenizer = tiktoken.get_encoding("gpt2")
    encoded = tokenizer.encode(start_context)
    encoded_tensor = torch.tensor(encoded, device=device).unsqueeze(0)

    print(f"\n{50*'='}\n{22*' '}IN\n{50*'='}")
    print("\nInput text:", start_context)
    print("Encoded input text:", encoded)
    print("encoded_tensor.shape:", encoded_tensor.shape)

    if torch.cuda.is_available():
        torch.cuda.synchronize()
    start = time.time()

    token_ids = generate_text_simple(
        model=model,
        idx=encoded_tensor,
        max_new_tokens=1000,
        context_size=GPT_CONFIG_124M["context_length"]
    )
    if torch.cuda.is_available():
        torch.cuda.synchronize()
    total_time = time.time() - start

    decoded_text = tokenizer.decode(token_ids.squeeze(0).tolist())

    print(f"\n\n{50*'='}\n{22*' '}OUT\n{50*'='}")
    print("\nOutput:", token_ids)
    print("Output length:", len(token_ids[0]))
    # print("Output text:", decoded_text)

    print(f"\nTime: {total_time:.2f} sec")
    print(f"{int(len(token_ids[0])/total_time)} tokens/sec")
    if torch.cuda.is_available():
        max_mem_bytes = torch.cuda.max_memory_allocated()
        max_mem_gb = max_mem_bytes / (1024 ** 3)
        print(f"Max memory allocated: {max_mem_gb:.2f} GB")


if __name__ == "__main__":
    main()
相关推荐
码云骑士1 小时前
116-Python调用GPT-4V分析图片-Base64编码-多图-JSON结构化输出
开发语言·python·json
程序员cxuan1 小时前
一招让你的 Claude 被封了也能用
人工智能·后端·程序员
赛逸会展s1 小时前
2027北京AI健康科技与智慧医疗展6月举办：不是秀场是成交现场
人工智能·科技
Hello_Damon_Nikola1 小时前
TPA3116D2DADR从入门到精通
人工智能·单片机·嵌入式硬件·物联网
Dachui_11221 小时前
飞牛 NAS 部署 Hermes AI Agent + ZeroNews 内网穿透，实现远程可管理的 AI 助手
人工智能·ai·远程工作·内网穿透
老郑聊AI业财智造1 小时前
TensorFlow 技术架构与源码分析
人工智能·python·深度学习·架构·tensorflow·软件工程
民乐团扒谱机1 小时前
【微科普】压缩感知（CS）：违反了奈奎斯特采样定理？不先采集也能还原信号？大白话讲透稀疏采样、L1重建与OMP，一文吃透附代码
python·神经网络·线性代数·算法·数学建模·压缩感知·奈奎斯特
海云安1 小时前
AI安全体系化治理：管理制度、技术管控与运营评估如何闭环
人工智能·安全
m4Rk_1 小时前
【论文阅读】Agent 记忆机制（49）：MemCoE——先学习如何组织记忆，再学习具体记住什么
论文阅读·人工智能·学习·开源·github
热门推荐
01国内可直接用、免费额度/永久免费的大模型API清单（含 SiliconFlow、火山、阿里、智谱、百度、Kimi、DeepSeek、DMXAPI 等）02GitHub 镜像站点032026 国产 AI 大模型横评：DeepSeek、通义千问、Kimi、文心一言、星火、豆包谁更能打？04微信历史版本含下载地址（ Windows PC | 安卓 | MAC ）及设置微信不更新05历年考研数学一、数学二、数学三真题试卷及答案PDF06Agnes AI 免费 API 接入指南：文本、生图、生视频，一套接口全免费07AI 编程 IDE 全景解析 2026：Agent 全面接管开发链路082026年7月AI圈大地震：GPT-5.6被政府限制、Claude入驻Slack、Anthropic自研芯片09VSCode+CodeX扩展，无法启动10如何新建文件夹？ 电脑新建文件夹的4种方法