Model

"""
Full definition of a GPT Language Model, all of it in this single file.
References:
1) the official GPT-2 TensorFlow implementation released by OpenAI:
https://github.com/openai/gpt-2/blob/master/src/model.py
2) huggingface/transformers PyTorch implementation:
https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py
"""

import math
import inspect
from dataclasses import dataclass

import torch
import torch.nn as nn
from torch.nn import functional as F

"""
PyTorch 的内置层归一化并不直接支持禁用偏置，所以自定义一个层归一化类。

self.weight: 创建一个可学习的参数 weight，初始值为全1的张量，维度为 nd。
self.bias: 如果 bias 为真，则创建一个可学习的参数 bias，初始值为全0的张量，维度为 ndim；否则，self.bias 为 None。

F.layer_norm 是 PyTorch 提供的函数，用于执行层归一化。
input: 输入张量。
self.weight.shape: 权重的形状，确保在归一化过程中使用正确的权重。
self.weight: 权重参数。
self.bias: 偏置参数（如果存在）。
1e-5: 归一化时添加的极小值，防止分母为零。
"""
class LayerNorm(nn.Module):
    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """

    def __init__(self, ndim, bias):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None

    def forward(self, input):
        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)


"""
self.c_attn：用于计算查询（query）、键（key）和值（value）的线性变换，输出维度为 3 * n_embd。
self.c_proj：用于将注意力输出进行线性变换，输出维度为 n_embd。

register_buffer 是 nn.Module 中的一个方法，用于将一个张量注册为模型的缓冲区。这些缓冲区在模型保存和加载时会被保留，但不会被视为模型的可学习参数（即不会参与梯度计算）。
在这里，"bias" 是缓冲区的名称，可以通过 self.bias 访问。
这个因果掩码（bias）确保在进行自注意力计算时，每个位置只能看到自己及之前的输入。这是实现因果自注意力机制的关键，避免模型在生成过程中看到未来的信息。
使用 register_buffer 的好处是，这个掩码在模型保存和加载时会被正确处理，而不需要将其作为可学习参数管理。

B: 批次大小（batch size）
T: 序列长度（sequence length）
C: 嵌入维度（n_embd）

将 k、q 和 v 的形状调整为适合多头注意力的格式：(B, n_head, T, hs)，其中 hs 是每个头的隐藏层维度。

scaled_dot_product_attention：
query: Tensor：查询矩阵，通常表示要关注的输入部分。
key: Tensor：键矩阵，表示用于匹配的输入部分。
value: Tensor：值矩阵，表示实际的信息来源。
attn_mask: Optional[Tensor]：可选的注意力掩码，通常用于遮盖某些位置（例如，填充标记或未来的信息）。
dropout_p: float：在计算注意力后应用的 dropout 概率，用于正则化。
is_causal: bool：指示是否使用因果注意力。如果为真，表示当前时间步只能关注之前的时间步。
scale: Optional[float]：缩放因子，通常是键的维度的平方根，用于避免点积过大。

dim=-1 表示在最后一个维度上进行 softmax 操作。在这个上下文中，dim=-1 意味着对注意力分数的每一行（对应于一个查询）应用 softmax。
这意味着对于每个查询位置，softmax 将所有键的位置的注意力分数转换为权重，权重和为1。
"""
class CausalSelfAttention(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        # key, query, value projections for all heads, but in a batch
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
        # output projection
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
        # regularization
        self.attn_dropout = nn.Dropout(config.dropout)
        self.resid_dropout = nn.Dropout(config.dropout)
        self.n_head = config.n_head
        self.n_embd = config.n_embd
        self.dropout = config.dropout
        # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0
        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
        if not self.flash:
            print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
            # causal mask to ensure that attention is only applied to the left in the input sequence
            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
                                        .view(1, 1, config.block_size, config.block_size))

    def forward(self, x):
        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)

        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
        q, k, v  = self.c_attn(x).split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)

        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
        if self.flash:
            # efficient attention using Flash Attention CUDA kernels
            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout if self.training else 0, is_causal=True)
        else:
            # manual implementation of attention
            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
            att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
            att = F.softmax(att, dim=-1)
            att = self.attn_dropout(att)
            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side

        # output projection
        y = self.resid_dropout(self.c_proj(y))
        return y


"""
使用 GELU（Gaussian Error Linear Unit）作为激活函数。这是一种流行的非线性激活函数，通常用于现代神经网络中，尤其是 Transformer 相关模型。
GELU 激活函数的输出会对输入进行加权，结合了输入的值和其在标准正态分布下的概率。
这使得负值会被压制（接近于零），而正值则会得到增强，特别是值越大时。
"""
class MLP(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
        self.gelu    = nn.GELU()
        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x

"""
这里先做了LayerNorm，再进行计算。
"""
class Block(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
        self.mlp = MLP(config)

    def forward(self, x):
        x = x + self.attn(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x

@dataclass
class GPTConfig:
    block_size: int = 1024
    vocab_size: int = 50304 # GPT-2 vocab_size of 50257, padded up to nearest multiple of 64 for efficiency
    n_layer: int = 12
    n_head: int = 12
    n_embd: int = 768
    dropout: float = 0.0
    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster



"""
wte：词嵌入层，用于将词索引映射到嵌入空间。
wpe：位置嵌入层，用于编码单词在序列中的位置。
drop：Dropout 层，用于防止过拟合。
h：多个 Transformer 块（Block）的列表。
ln_f：最终层的层归一化。

self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
输出层，将嵌入转换为词汇表的 logits。

self.transformer.wte.weight = self.lm_head.weight
使用权重共享，lm_head 的权重与 wte 的权重相同，以减少参数量。

通过迭代模型的所有参数，使用 p.numel() 计算每个参数的元素总数，并将它们相加，得到总参数数量 n_params。
默认情况下，如果 non_embedding 为 True，则不计算位置嵌入的参数。

根据模块类型初始化权重：
对于 nn.Linear 和 nn.Embedding，使用正态分布初始化。

如果提供了目标，则计算 logits 和损失；否则只计算最后一个位置的 logits。

允许在需要时调整块大小，适用于加载预训练模型时。

weight_decay: 权重衰减参数，用于控制 L2 正则化。
"""
class GPT(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.vocab_size is not None
        assert config.block_size is not None
        self.config = config

        self.transformer = nn.ModuleDict(dict(
            wte = nn.Embedding(config.vocab_size, config.n_embd),
            wpe = nn.Embedding(config.block_size, config.n_embd),
            drop = nn.Dropout(config.dropout),
            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
            ln_f = LayerNorm(config.n_embd, bias=config.bias),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        # with weight tying when using torch.compile() some warnings get generated:
        # "UserWarning: functional_call was passed multiple values for tied weights.
        # This behavior is deprecated and will be an error in future versions"
        # not 100% sure what this is, so far seems to be harmless. TODO investigate
        self.transformer.wte.weight = self.lm_head.weight # https://paperswithcode.com/method/weight-tying

        # init all weights
        self.apply(self._init_weights)
        # apply special scaled init to the residual projections, per GPT-2 paper
        for pn, p in self.named_parameters():
            if pn.endswith('c_proj.weight'):
                torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))

        # report number of parameters
        print("number of parameters: %.2fM" % (self.get_num_params()/1e6,))

    def get_num_params(self, non_embedding=True):
        """
        Return the number of parameters in the model.
        For non-embedding count (default), the position embeddings get subtracted.
        The token embeddings would too, except due to the parameter sharing these
        params are actually used as weights in the final layer, so we include them.
        """
        n_params = sum(p.numel() for p in self.parameters())
        if non_embedding:
            n_params -= self.transformer.wpe.weight.numel()
        return n_params

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, idx, targets=None):
        device = idx.device
        b, t = idx.size()
        assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
        pos = torch.arange(0, t, dtype=torch.long, device=device) # shape (t)

        # forward the GPT model itself
        tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (t, n_embd)
        x = self.transformer.drop(tok_emb + pos_emb)
        for block in self.transformer.h:
            x = block(x)
        x = self.transformer.ln_f(x)

        if targets is not None:
            # if we are given some desired targets also calculate the loss
            logits = self.lm_head(x)
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
        else:
            # inference-time mini-optimization: only forward the lm_head on the very last position
            logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
            loss = None

        return logits, loss

    def crop_block_size(self, block_size):
        # model surgery to decrease the block size if necessary
        # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024)
        # but want to use a smaller block size for some smaller, simpler model
        assert block_size <= self.config.block_size
        self.config.block_size = block_size
        self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:block_size])
        for block in self.transformer.h:
            if hasattr(block.attn, 'bias'):
                block.attn.bias = block.attn.bias[:,:,:block_size,:block_size]

    @classmethod
    def from_pretrained(cls, model_type, override_args=None):
        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
        override_args = override_args or {} # default to empty dict
        # only dropout can be overridden see more notes below
        assert all(k == 'dropout' for k in override_args)
        from transformers import GPT2LMHeadModel
        print("loading weights from pretrained gpt: %s" % model_type)

        # n_layer, n_head and n_embd are determined from model_type
        config_args = {
            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
        }[model_type]
        print("forcing vocab_size=50257, block_size=1024, bias=True")
        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
        config_args['bias'] = True # always True for GPT model checkpoints
        # we can override the dropout rate, if desired
        if 'dropout' in override_args:
            print(f"overriding dropout rate to {override_args['dropout']}")
            config_args['dropout'] = override_args['dropout']
        # create a from-scratch initialized minGPT model
        config = GPTConfig(**config_args)
        model = GPT(config)
        sd = model.state_dict()
        sd_keys = sd.keys()
        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param

        # init a huggingface/transformers model
        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
        sd_hf = model_hf.state_dict()

        # copy while ensuring all of the parameters are aligned and match in names and shapes
        sd_keys_hf = sd_hf.keys()
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
        # this means that we have to transpose these weights when we import them
        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
        for k in sd_keys_hf:
            if any(k.endswith(w) for w in transposed):
                # special treatment for the Conv1D weights we need to transpose
                assert sd_hf[k].shape[::-1] == sd[k].shape
                with torch.no_grad():
                    sd[k].copy_(sd_hf[k].t())
            else:
                # vanilla copy over the other parameters
                assert sd_hf[k].shape == sd[k].shape
                with torch.no_grad():
                    sd[k].copy_(sd_hf[k])

        return model

    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
        # start with all of the candidate parameters
        param_dict = {pn: p for pn, p in self.named_parameters()}
        # filter out those that do not require grad
        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
        optim_groups = [
            {'params': decay_params, 'weight_decay': weight_decay},
            {'params': nodecay_params, 'weight_decay': 0.0}
        ]
        num_decay_params = sum(p.numel() for p in decay_params)
        num_nodecay_params = sum(p.numel() for p in nodecay_params)
        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        # Create AdamW optimizer and use the fused version if it is available
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == 'cuda'
        extra_args = dict(fused=True) if use_fused else dict()
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args)
        print(f"using fused AdamW: {use_fused}")

        return optimizer

    def estimate_mfu(self, fwdbwd_per_iter, dt):
        """ estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """
        # first estimate the number of flops we do per iteration.
        # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311
        N = self.get_num_params()
        cfg = self.config
        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd//cfg.n_head, cfg.block_size
        flops_per_token = 6*N + 12*L*H*Q*T
        flops_per_fwdbwd = flops_per_token * T
        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
        # express our flops throughput as ratio of A100 bfloat16 peak flops
        flops_achieved = flops_per_iter * (1.0/dt) # per second
        flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS
        mfu = flops_achieved / flops_promised
        return mfu

    @torch.no_grad()
    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
        """
        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
        the sequence max_new_tokens times, feeding the predictions back into the model each time.
        Most likely you'll want to make sure to be in model.eval() mode of operation for this.
        """
        for _ in range(max_new_tokens):
            # if the sequence context is growing too long we must crop it at block_size
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            # forward the model to get the logits for the index in the sequence
            logits, _ = self(idx_cond)
            # pluck the logits at the final step and scale by desired temperature
            logits = logits[:, -1, :] / temperature
            # optionally crop the logits to only the top k options
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = -float('Inf')
            # apply softmax to convert logits to (normalized) probabilities
            probs = F.softmax(logits, dim=-1)
            # sample from the distribution
            idx_next = torch.multinomial(probs, num_samples=1)
            # append sampled index to the running sequence and continue
            idx = torch.cat((idx, idx_next), dim=1)

        return idx

Train

"""
This training script can be run both on a single gpu in debug mode,
and also in a larger training run with distributed data parallel (ddp).

To run on a single GPU, example:
$ python train.py --batch_size=32 --compile=False

To run with DDP on 4 gpus on 1 node, example:
$ torchrun --standalone --nproc_per_node=4 train.py

To run with DDP on 4 gpus across 2 nodes, example:
- Run on the first (master) node with example IP 123.456.123.456:
$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr=123.456.123.456 --master_port=1234 train.py
- Run on the worker node:
$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr=123.456.123.456 --master_port=1234 train.py
(If your cluster does not have Infiniband interconnect prepend NCCL_IB_DISABLE=1)
"""

import os
import time
import math
import pickle
from contextlib import nullcontext

import numpy as np
import torch
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group

from model import GPTConfig, GPT

# -----------------------------------------------------------------------------
# default config values designed to train a gpt2 (124M) on OpenWebText
# I/O
out_dir = 'out'
eval_interval = 2000
log_interval = 1
eval_iters = 200
eval_only = False # if True, script exits right after the first eval
always_save_checkpoint = True # if True, always save a checkpoint after each eval
init_from = 'scratch' # 'scratch' or 'resume' or 'gpt2*'
# wandb logging
wandb_log = False # disabled by default
wandb_project = 'owt'
wandb_run_name = 'gpt2' # 'run' + str(time.time())
# data
dataset = 'openwebtext'
"""
gradient_accumulation_steps: 梯度累积的步数。在训练过程中，通常会在多个小批次上累积梯度，以减少显存占用并模拟较大的批次大小。
gradient_accumulation_steps 决定了在更新梯度之前会处理多少个小批次的数据。
ddp_world_size 表示并行训练时，所有 GPU 一共会处理多少数据。
batch_size 和 block_size 则决定了每个批次中有多少 token 被处理。
最终，这个计算给出了在一个训练迭代中实际处理的 token 总数，这有助于评估模型的训练负载和效率。
"""
gradient_accumulation_steps = 5 * 8 # used to simulate larger batch sizes
batch_size = 12 # if gradient_accumulation_steps > 1, this is the micro-batch size
block_size = 1024
# model
n_layer = 12
n_head = 12
n_embd = 768
dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+
bias = False # do we use bias inside LayerNorm and Linear layers?
# adamw optimizer
learning_rate = 6e-4 # max learning rate
max_iters = 600000 # total number of training iterations
weight_decay = 1e-1
beta1 = 0.9
beta2 = 0.95
grad_clip = 1.0 # clip gradients at this value, or disable if == 0.0
# learning rate decay settings
decay_lr = True # whether to decay the learning rate
warmup_iters = 2000 # how many steps to warm up for
lr_decay_iters = 600000 # should be ~= max_iters per Chinchilla
min_lr = 6e-5 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla
# DDP settings
backend = 'nccl' # 'nccl', 'gloo', etc.
# system
device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks
dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler
compile = True # use PyTorch 2.0 to compile the model to be faster
# -----------------------------------------------------------------------------
config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open('configurator.py').read()) # overrides from command line or config file
config = {k: globals()[k] for k in config_keys} # will be useful for logging
# -----------------------------------------------------------------------------

# various inits, derived attributes, I/O setup
ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?
if ddp:
    init_process_group(backend=backend)
    ddp_rank = int(os.environ['RANK'])
    ddp_local_rank = int(os.environ['LOCAL_RANK'])
    ddp_world_size = int(os.environ['WORLD_SIZE'])
    device = f'cuda:{ddp_local_rank}'
    torch.cuda.set_device(device)
    master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.
    seed_offset = ddp_rank # each process gets a different seed
    # world_size number of processes will be training simultaneously, so we can scale
    # down the desired gradient accumulation iterations per process proportionally
    assert gradient_accumulation_steps % ddp_world_size == 0
    gradient_accumulation_steps //= ddp_world_size
else:
    # if not ddp, we are running on a single gpu, and one process
    master_process = True
    seed_offset = 0
    ddp_world_size = 1
tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
print(f"tokens per iteration will be: {tokens_per_iter:,}")

if master_process:
    os.makedirs(out_dir, exist_ok=True)
torch.manual_seed(1337 + seed_offset)
"""
torch.backends.cuda.matmul.allow_tf32 = True
允许在 CUDA 的矩阵乘法运算中使用 TensorFloat-32（TF32）。TF32 是一种在 NVIDIA Ampere 架构及以上 GPU 上优化浮点运算的格式，能够在保持计算精度的同时提升性能。

torch.backends.cudnn.allow_tf32 = True
允许在 cuDNN 中使用 TF32。这同样是为了提升深度学习模型在使用 NVIDIA GPU 时的计算效率。

device_type = 'cuda' if 'cuda' in device else 'cpu'
根据当前设备的类型来确定 device_type。如果设备是 CUDA（即使用 GPU），则设置为 'cuda'，否则设置为 'cpu'。

ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
根据传入的 dtype（数据类型）来设置 PyTorch 的数据类型。支持的类型包括：

'float32'：单精度浮点数。
'bfloat16'：BFLOAT16 格式，适合用于深度学习训练。
'float16'：半精度浮点数。
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
设置上下文管理器 ctx：

如果设备是 CPU，则使用 nullcontext()，表示不进行任何特别的上下文管理。
如果设备是 GPU，使用 torch.amp.autocast，它会自动管理混合精度训练，优化内存使用并加速计算。
"""
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
# note: float16 data type will automatically use a GradScaler
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

# poor man's data loader
data_dir = os.path.join('data', dataset)
"""
如果是训练集，使用 numpy.memmap 加载训练数据文件 train.bin。np.memmap 允许大数据集的部分加载，有助于节省内存。

生成 batch_size 个随机起始索引，范围在数据长度减去 block_size 之间。这个确保每个批次的序列不会超出数据的界限。

使用生成的索引 ix 创建输入批次 x。对于每个索引 i，提取 block_size 长度的序列，将其转换为 PyTorch 张量并堆叠成一个批次。

如果是 GPU，首先将张量 x 和 y 锁定到固定内存中（pin_memory()），这样可以异步地将它们传输到 GPU（non_blocking=True），提高效率。
"""
def get_batch(split):
    # We recreate np.memmap every batch to avoid a memory leak, as per
    # https://stackoverflow.com/questions/45132940/numpy-memmap-memory-usage-want-to-iterate-once/61472122#61472122
    if split == 'train':
        data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
    else:
        data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')
    ix = torch.randint(len(data) - block_size, (batch_size,))
    x = torch.stack([torch.from_numpy((data[i:i+block_size]).astype(np.int64)) for i in ix])
    y = torch.stack([torch.from_numpy((data[i+1:i+1+block_size]).astype(np.int64)) for i in ix])
    if device_type == 'cuda':
        # pin arrays x,y, which allows us to move them to GPU asynchronously (non_blocking=True)
        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
    else:
        x, y = x.to(device), y.to(device)
    return x, y

# init these up here, can override if init_from='resume' (i.e. from a checkpoint)
iter_num = 0
best_val_loss = 1e9

# attempt to derive vocab_size from the dataset
meta_path = os.path.join(data_dir, 'meta.pkl')
meta_vocab_size = None
if os.path.exists(meta_path):
    with open(meta_path, 'rb') as f:
        meta = pickle.load(f)
    meta_vocab_size = meta['vocab_size']
    print(f"found vocab_size = {meta_vocab_size} (inside {meta_path})")

# model init
model_args = dict(n_layer=n_layer, n_head=n_head, n_embd=n_embd, block_size=block_size,
                  bias=bias, vocab_size=None, dropout=dropout) # start with model_args from command line
"""
使用 map_location=device 确保将模型加载到正确的设备上（CPU 或 GPU）。

对一些关键参数进行强制更新，以确保在恢复时与保存时的模型配置一致。这些参数包括层数、头数、嵌入维度、块大小、是否使用偏置和词汇表大小。

这行代码初始化一个 GradScaler，用于在使用混合精度训练时缩放梯度。只有当 dtype 为 float16 时，梯度缩放器才会启用。混合精度训练可以提高训练效率并减少内存使用。
"""
if init_from == 'scratch':
    # init a new model from scratch
    print("Initializing a new model from scratch")
    # determine the vocab size we'll use for from-scratch training
    if meta_vocab_size is None:
        print("defaulting to vocab_size of GPT-2 to 50304 (50257 rounded up for efficiency)")
    model_args['vocab_size'] = meta_vocab_size if meta_vocab_size is not None else 50304
    gptconf = GPTConfig(**model_args)
    model = GPT(gptconf)
elif init_from == 'resume':
    print(f"Resuming training from {out_dir}")
    # resume training from a checkpoint.
    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device)
    checkpoint_model_args = checkpoint['model_args']
    # force these config attributes to be equal otherwise we can't even resume training
    # the rest of the attributes (e.g. dropout) can stay as desired from command line
    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
        model_args[k] = checkpoint_model_args[k]
    # create the model
    gptconf = GPTConfig(**model_args)
    model = GPT(gptconf)
    state_dict = checkpoint['model']
    # fix the keys of the state dictionary :(
    # honestly no idea how checkpoints sometimes get this prefix, have to debug more
    unwanted_prefix = '_orig_mod.'
    for k,v in list(state_dict.items()):
        if k.startswith(unwanted_prefix):
            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
    model.load_state_dict(state_dict)
    iter_num = checkpoint['iter_num']
    best_val_loss = checkpoint['best_val_loss']
elif init_from.startswith('gpt2'):
    print(f"Initializing from OpenAI GPT-2 weights: {init_from}")
    # initialize from OpenAI GPT-2 weights
    override_args = dict(dropout=dropout)
    model = GPT.from_pretrained(init_from, override_args)
    # read off the created config params, so we can store them into checkpoint correctly
    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
        model_args[k] = getattr(model.config, k)
# crop down the model block size if desired, using model surgery
if block_size < model.config.block_size:
    model.crop_block_size(block_size)
    model_args['block_size'] = block_size # so that the checkpoint will have the right value
model.to(device)

# initialize a GradScaler. If enabled=False scaler is a no-op
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))

# optimizer
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
if init_from == 'resume':
    optimizer.load_state_dict(checkpoint['optimizer'])
checkpoint = None # free up memory

# compile the model
if compile:
    print("compiling the model... (takes a ~minute)")
    unoptimized_model = model
    model = torch.compile(model) # requires PyTorch 2.0

# wrap model into DDP container
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])

# helps estimate an arbitrarily accurate loss over either split using many batches
@torch.no_grad()
def estimate_loss():
    out = {}
    model.eval()
    for split in ['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            with ctx:
                logits, loss = model(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    model.train()
    return out

# learning rate decay scheduler (cosine with warmup)
def get_lr(it):
    # 1) linear warmup for warmup_iters steps
    if it < warmup_iters:
        return learning_rate * it / warmup_iters
    # 2) if it > lr_decay_iters, return min learning rate
    if it > lr_decay_iters:
        return min_lr
    # 3) in between, use cosine decay down to min learning rate
    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
    assert 0 <= decay_ratio <= 1
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
    return min_lr + coeff * (learning_rate - min_lr)

# logging
if wandb_log and master_process:
    import wandb
    wandb.init(project=wandb_project, name=wandb_run_name, config=config)

# training loop
X, Y = get_batch('train') # fetch the very first batch
t0 = time.time()
local_iter_num = 0 # number of iterations in the lifetime of this process
raw_model = model.module if ddp else model # unwrap DDP container if needed
running_mfu = -1.0
while True:

    # determine and set the learning rate for this iteration
    lr = get_lr(iter_num) if decay_lr else learning_rate
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    # evaluate the loss on train/val sets and write checkpoints
    if iter_num % eval_interval == 0 and master_process:
        losses = estimate_loss()
        print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
        if wandb_log:
            wandb.log({
                "iter": iter_num,
                "train/loss": losses['train'],
                "val/loss": losses['val'],
                "lr": lr,
                "mfu": running_mfu*100, # convert to percentage
            })
        if losses['val'] < best_val_loss or always_save_checkpoint:
            best_val_loss = losses['val']
            if iter_num > 0:
                checkpoint = {
                    'model': raw_model.state_dict(),
                    'optimizer': optimizer.state_dict(),
                    'model_args': model_args,
                    'iter_num': iter_num,
                    'best_val_loss': best_val_loss,
                    'config': config,
                }
                print(f"saving checkpoint to {out_dir}")
                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))
    if iter_num == 0 and eval_only:
        break

    # forward backward update, with optional gradient accumulation to simulate larger batch size
    # and using the GradScaler if data type is float16
    for micro_step in range(gradient_accumulation_steps):
        if ddp:
            # in DDP training we only need to sync gradients at the last micro step.
            # the official way to do this is with model.no_sync() context manager, but
            # I really dislike that this bloats the code and forces us to repeat code
            # looking at the source of that context manager, it just toggles this variable
            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
        with ctx:
            logits, loss = model(X, Y)
            loss = loss / gradient_accumulation_steps # scale the loss to account for gradient accumulation
        # immediately async prefetch next batch while model is doing the forward pass on the GPU
        X, Y = get_batch('train')
        # backward pass, with gradient scaling if training in fp16
        """使用 GradScaler 进行梯度缩放，确保在使用半精度浮点数时梯度不会下溢。"""
        scaler.scale(loss).backward()
    # clip the gradient
    """如果设置了梯度裁剪，则取消缩放并裁剪模型参数的梯度，防止梯度爆炸。"""
    if grad_clip != 0.0:
        scaler.unscale_(optimizer)
        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
    # step the optimizer and scaler if training in fp16
    scaler.step(optimizer)
    scaler.update()
    # flush the gradients as soon as we can, no need for this memory anymore
    optimizer.zero_grad(set_to_none=True)

    # timing and logging
    t1 = time.time()
    dt = t1 - t0
    t0 = t1
    if iter_num % log_interval == 0 and master_process:
        # get loss as float. note: this is a CPU-GPU sync point
        # scale up to undo the division above, approximating the true total loss (exact would have been a sum)
        lossf = loss.item() * gradient_accumulation_steps
        if local_iter_num >= 5: # let the training loop settle a bit
            mfu = raw_model.estimate_mfu(batch_size * gradient_accumulation_steps, dt)
            running_mfu = mfu if running_mfu == -1.0 else 0.9*running_mfu + 0.1*mfu
        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%")
    iter_num += 1
    local_iter_num += 1

    # termination conditions
    if iter_num > max_iters:
        break

if ddp:
    destroy_process_group()

Sample

"""
Sample from a trained model
"""
import os
import pickle
from contextlib import nullcontext
import torch
import tiktoken
from model import GPTConfig, GPT

# -----------------------------------------------------------------------------
init_from = 'resume' # either 'resume' (from an out_dir) or a gpt2 variant (e.g. 'gpt2-xl')
out_dir = 'out' # ignored if init_from is not 'resume'
start = "\n" # or "<|endoftext|>" or etc. Can also specify a file, use as: "FILE:prompt.txt"
num_samples = 10 # number of samples to draw
max_new_tokens = 500 # number of tokens generated in each sample
temperature = 0.8 # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions
top_k = 200 # retain only the top_k most likely tokens, clamp others to have 0 probability
seed = 1337
device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.
dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32' or 'bfloat16' or 'float16'
compile = False # use PyTorch 2.0 to compile the model to be faster
exec(open('configurator.py').read()) # overrides from command line or config file
# -----------------------------------------------------------------------------

torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

# model
if init_from == 'resume':
    # init from a model saved in a specific directory
    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device)
    gptconf = GPTConfig(**checkpoint['model_args'])
    model = GPT(gptconf)
    state_dict = checkpoint['model']
    unwanted_prefix = '_orig_mod.'
    for k,v in list(state_dict.items()):
        if k.startswith(unwanted_prefix):
            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
    model.load_state_dict(state_dict)
elif init_from.startswith('gpt2'):
    # init from a given GPT-2 model
    model = GPT.from_pretrained(init_from, dict(dropout=0.0))

model.eval()
model.to(device)
if compile:
    model = torch.compile(model) # requires PyTorch 2.0 (optional)

# look for the meta pickle in case it is available in the dataset folder
load_meta = False
if init_from == 'resume' and 'config' in checkpoint and 'dataset' in checkpoint['config']: # older checkpoints might not have these...
    meta_path = os.path.join('data', checkpoint['config']['dataset'], 'meta.pkl')
    load_meta = os.path.exists(meta_path)
if load_meta:
    print(f"Loading meta from {meta_path}...")
    with open(meta_path, 'rb') as f:
        meta = pickle.load(f)
    # TODO want to make this more general to arbitrary encoder/decoder schemes
    stoi, itos = meta['stoi'], meta['itos']
    encode = lambda s: [stoi[c] for c in s]
    decode = lambda l: ''.join([itos[i] for i in l])
else:
    # ok let's assume gpt-2 encodings by default
    print("No meta.pkl found, assuming GPT-2 encodings...")
    enc = tiktoken.get_encoding("gpt2")
    encode = lambda s: enc.encode(s, allowed_special={"<|endoftext|>"})
    decode = lambda l: enc.decode(l)

# encode the beginning of the prompt
if start.startswith('FILE:'):
    with open(start[5:], 'r', encoding='utf-8') as f:
        start = f.read()
start_ids = encode(start)
x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])

# run generation
with torch.no_grad():
    with ctx:
        for k in range(num_samples):
            y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
            print(decode(y[0].tolist()))
            print('---------------')