Upload original AdamW and Muon pretrained checkpoints (step 31000)
ad93bc0 verified 1 day ago
39.3 kB
"""
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
"""

"""
change the configure_optimizers method to handle combined optimizers
add activation hooks to enable activation preconditioning
"""

import math
import inspect
from dataclasses import dataclass

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



from optimizer.optimizer_strategies import OPTIMIZER_STRATEGIES, OptimizerConfig


# +++ NEW: RoPE Helper Functions +++
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
    """
    预计算 Cos 和 Sin 表。
    dim: head_dim (n_embd // n_head)
    end: block_size
    """
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
    t = torch.arange(end, device=freqs.device)  # type: ignore
    freqs = torch.outer(t, freqs).float()  # (end, dim // 2)
    freqs_cos = torch.cos(freqs)  # Real part
    freqs_sin = torch.sin(freqs)  # Imaginary part
    return freqs_cos, freqs_sin

def apply_rotary_emb(xq, xk, freqs_cos, freqs_sin):
    """
    应用旋转。
    xq, xk shape: (B, T, n_head, head_dim)
    freqs_cos, freqs_sin shape: (T, head_dim/2) -> 广播到 (1, T, 1, head_dim/2)
    """
    # reshape xq, xk to manipulate pairs
    # input: (B, T, n_head, head_dim)
    # 我们将 head_dim 拆分为 (head_dim//2, 2) 以便进行旋转
    
    # 将 freqs 切片以匹配当前序列长度 (用于推理时的 extrapolation)
    seq_len = xq.shape[1]
    freqs_cos = freqs_cos[:seq_len].view(1, seq_len, 1, -1)
    freqs_sin = freqs_sin[:seq_len].view(1, seq_len, 1, -1)

    # 核心旋转逻辑
    # x: [x1, x2, x3, x4] -> x_real: [x1, x3], x_imag: [x2, x4] (这是一种简单的 split 方式)
    # 这里我们采用 Llama/PaLM 常用的 split 方式：将前半部分作为 real，后半部分作为 imag
    d = xq.shape[-1] // 2
    
    xq_r, xq_i = xq[..., :d], xq[..., d:]
    xk_r, xk_i = xk[..., :d], xk[..., d:]
    
    # 旋转公式:
    # x' = x_r * cos - x_i * sin
    # y' = x_r * sin + x_i * cos
    
    xq_out_r = xq_r * freqs_cos - xq_i * freqs_sin
    xq_out_i = xq_r * freqs_sin + xq_i * freqs_cos
    
    xk_out_r = xk_r * freqs_cos - xk_i * freqs_sin
    xk_out_i = xk_r * freqs_sin + xk_i * freqs_cos
    
    # 拼接回去
    xq_out = torch.cat([xq_out_r, xq_out_i], dim=-1)
    xk_out = torch.cat([xk_out_r, xk_out_i], dim=-1)
    
    return xq_out, xk_out
# +++ END NEW +++


# +++ NEW: LoRA Layer Definition +++
class LoRALinear(nn.Module):
    """
    Implements a LoRA-enabled Linear layer.
    
    This layer contains a frozen, pre-trained linear layer and two 
    low-rank matrices (A and B) that are trainable.
    
    The forward pass computes: y = Wx + (s * BA)x
    where W is the frozen weight, B and A are the low-rank matrices,
    and s is the scaling factor (lora_alpha / r).
    """
    def __init__(self, in_features, out_features, r, lora_alpha, lora_dropout, bias=True):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.r = r
        self.lora_alpha = lora_alpha

        # 1. Original, frozen linear layer
        self.linear = nn.Linear(in_features, out_features, bias=bias)
        
        # 2. LoRA matrices (A and B)
        # B is (out, r) and A is (r, in)
        self.lora_A = nn.Parameter(torch.zeros(r, in_features))
        self.lora_B = nn.Parameter(torch.zeros(out_features, r))
        
        # 3. Scaling factor
        self.scaling = self.lora_alpha / self.r
        
        # 4. Dropout
        self.lora_dropout = nn.Dropout(lora_dropout)
        
        # 5. Freeze the original linear layer
        self.linear.requires_grad_(False)
        
        # 6. Initialize LoRA weights
        # Kaiming uniform for A, zeros for B
        # This ensures that (s * BA)x is zero at the beginning of training
        nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
        # nn.init.zeros_(self.lora_B) # lora_B is already initialized to zeros

    def forward(self, x: torch.Tensor):
        # Original, frozen path
        result = self.linear(x)
        
        # LoRA path
        # Note: We don't apply dropout to the original path (result)
        lora_update = self.lora_dropout(x)
        lora_update = lora_update @ self.lora_A.transpose(0, 1) # (B, T, in) @ (in, r) -> (B, T, r)
        lora_update = lora_update @ self.lora_B.transpose(0, 1) # (B, T, r) @ (r, out) -> (B, T, out)
        
        return result + lora_update * self.scaling

    def __repr__(self):
        return (f"LoRALinear(in_features={self.in_features}, out_features={self.out_features}, "
                f"r={self.r}, lora_alpha={self.lora_alpha}, bias={self.linear.bias is not None})")


# +++ MODIFIED: Updated LayerNorm implementation +++
class LayerNorm(nn.Module):
    """ LayerNorm that supports p-norms (1, 2, inf). """

    def __init__(self, ndim, bias, p=2.0):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
        self.p = p
        self.eps = 1e-5

    def forward(self, x):
        # x shape: (Batch, Time, Channels)
        
        # 1. Calculate Mean and Center
        mu = x.mean(dim=-1, keepdim=True)
        x_centered = x - mu

        # 2. Calculate Normalization Factor based on p
        if self.p == 2:
            # Standard LayerNorm (Scale by Standard Deviation)
            # using F.layer_norm is faster and more stable for p=2
            return F.layer_norm(x, self.weight.shape, self.weight, self.bias, self.eps)
        
        elif self.p == 1:
            # L1 Norm: Scale by Mean Absolute Deviation
            # norm = mean(|x - mu|)
            norm_factor = x_centered.abs().mean(dim=-1, keepdim=True)
            x_norm = x_centered / (norm_factor + self.eps)
            
        elif self.p == float('inf'):
            # L_inf Norm: Scale by Max Absolute Deviation
            # norm = max(|x - mu|)
            norm_factor = x_centered.abs().max(dim=-1, keepdim=True)[0]
            x_norm = x_centered / (norm_factor + self.eps)
            
        else:
            # Generic p-norm (average based)
            # norm = (mean(|x - mu|^p))^(1/p)
            norm_factor = x_centered.abs().pow(self.p).mean(dim=-1, keepdim=True).pow(1/self.p)
            x_norm = x_centered / (norm_factor + self.eps)

        # 3. Apply Affine Transform
        if self.bias is not None:
            return x_norm * self.weight + self.bias
        else:
            return x_norm * self.weight
# +++ END MODIFIED +++

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))

    # +++ MODIFIED: Forward 签名增加 RoPE 参数 +++
    def forward(self, x, freqs_cos=None, freqs_sin=None):
        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)

        # +++ NEW: Apply RoPE if provided +++
        if freqs_cos is not None and freqs_sin is not None:
            # RoPE expects (B, T, nh, hs), so we transpose back temporarily
            k = k.transpose(1, 2)
            q = q.transpose(1, 2)
            
            q, k = apply_rotary_emb(q, k, freqs_cos, freqs_sin)
            
            # Transpose back to (B, nh, T, hs) for attention calculation
            k = k.transpose(1, 2)
            q = q.transpose(1, 2)
        # +++ END NEW +++

        # 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

class MLP(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.activation_function = config.activation_function
        
        if self.activation_function == 'swiglu':
            # SwiGLU uses a different structure and hidden dimension, often 2/3 of the standard 4*n_embd
            # This is a common practice from papers like LLaMA
            hidden_dim = int(2 * (4 * config.n_embd) / 3)
            self.c_fc   = nn.Linear(config.n_embd, 2 * hidden_dim, bias=config.bias) # Projects to both gate and value
            self.c_proj = nn.Linear(hidden_dim, config.n_embd, bias=config.bias)
        
        elif self.activation_function == 'gelu':
            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)
        
        else:
            raise ValueError(f"Unsupported activation function: {self.activation_function}")
        
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        
        if self.activation_function == 'swiglu':
            # Project to get both gate and value tensors
            gate, value = self.c_fc(x).chunk(2, dim=-1)
            # Apply the SwiGLU activation: silu(gate) * value
            x = F.silu(gate) * value
        
        else: # 'gelu'
            x = self.c_fc(x)
            x = self.gelu(x)
        
        x = self.c_proj(x)
        x = self.dropout(x)
        return x

class Block(nn.Module):

    def __init__(self, config):
        super().__init__()
        # +++ MODIFIED: Pass layer_norm_p to LayerNorm +++
        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias, p=config.layer_norm_p)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias, p=config.layer_norm_p)
        # +++ END MODIFIED +++
        self.mlp = MLP(config)

    def forward(self, x, freqs_cos=None, freqs_sin=None):
        x = x + self.attn(self.ln_1(x), freqs_cos=freqs_cos, freqs_sin=freqs_sin)
        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
    
    # +++ NEW: Config for Positional Embedding +++
    pos_emb_type: str = 'absolute' # 'absolute' or 'rope'
    # +++ NEW: Layer Norm P +++
    layer_norm_p: float = 2.0
    # +++ END NEW +++
    
    # add use_activation_precond
    use_activation_precond: bool = False # whether to use activation preconditioning
    activation_function: str = 'gelu' # 'gelu' or 'swiglu'
    use_activation_monitoring: bool = False # whether to capture activations for monitoring

    # +++ NEW: Initialization Scale Parameters +++
    attn_init_scale: float = 0.02 # Base scale for QKV weights
    mlp_init_scale: float = 0.02 # Base scale for MLP expansion c_fc weights
    emb_init_scale: float = 0.02 # Base scale for Embedding weights
    # +++ END NEW +++

    # +++ NEW: LoRA Configuration +++
    use_lora: bool = False
    lora_r: int = 8
    lora_alpha: float = 16.0
    lora_dropout: float = 0.05
    # List of module name suffixes to apply LoRA to.
    # E.g., 'attn.c_attn' will target 'transformer.h.0.attn.c_attn'
    lora_target_modules: list = None
    # +++ END NEW +++

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),
            # +++ MODIFIED: Conditional WPE +++
            # 如果是 'rope'，我们不需要学习的 wpe，但为了兼容性或简单起见，可以设为 None 或 Identity
            # 为了严谨，如果是 RoPE，我们就不初始化 wpe
            wpe = nn.Embedding(config.block_size, config.n_embd) if config.pos_emb_type == 'absolute' else None,
            # +++ END MODIFIED +++
            drop = nn.Dropout(config.dropout),
            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
            # +++ MODIFIED: Pass p to final LayerNorm +++
            ln_f = LayerNorm(config.n_embd, bias=config.bias, p=config.layer_norm_p),
            # +++ END MODIFIED +++
        ))

        # +++ NEW: Initialize RoPE Cache +++
        if config.pos_emb_type == 'rope':
            # head_dim
            head_dim = config.n_embd // config.n_head
            # 预计算 cos/sin table，注册为 buffer (不作为参数更新，但随模型保存)
            cos, sin = precompute_freqs_cis(head_dim, config.block_size)
            self.register_buffer("freqs_cos", cos)
            self.register_buffer("freqs_sin", sin)
        # +++ END NEW +++

        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))
        
        # 2. 自定义初始化覆盖 (Custom Initialization Override)
        # 包含了 Q, K, V, MLP Expansion 以及 Residual Projections 的缩放
        self.custom_init_weights()

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

        # +++ NEW: Add a dictionary to store activations and register monitoring hooks +++
        if config.use_activation_monitoring:
            self.monitoring_activations = {}
            self.register_monitoring_hooks()

        # +++ ADDED: Register forward hooks for activation preconditioning +++
        # self.register_activation_hooks()
        if config.use_activation_precond:
            self.register_preconditioning_hooks()
    
    # NEW: custom initialization
    def custom_init_weights(self):
        """
        应用自定义的初始化标准差到 Attention 和 MLP 层。
        同时应用 GPT-2 论文中的残差缩放逻辑：std = base_std / sqrt(2 * n_layer)
        """
        config = self.config
        
        # 计算残差层的缩放因子
        # GPT-2 论文建议：随着层数增加，残差路径上的累积方差会增大，因此需要按层数缩放初始化值
        resid_scale_factor = 1.0 / math.sqrt(2.0 * config.n_layer)

        print(f"Applying custom init: attn_scale={config.attn_init_scale}, emb_scale={config.emb_init_scale}, mlp_scale={config.mlp_init_scale}, resid_factor={resid_scale_factor:.4f}")

        for name, module in self.named_modules():
            if isinstance(module, nn.Linear):
                target_std = None
                
                # --- Attention Block ---
                # 1. Q, K, V (Attention Input Projections)
                # use attn_init_scale
                if any(name.endswith(f'.attn.c_{t}') for t in ['q', 'k', 'v']) or name.endswith('.attn.c_attn'):
                    # note that we are changing here to allow c_attn
                    target_std = config.attn_init_scale
                
                # 2. Attention Output Projection (Residual Layer)
                # 使用 attn_init_scale * resid_scale_factor
                elif name.endswith('.attn.c_proj'):
                    target_std = config.attn_init_scale * resid_scale_factor

                # --- MLP Block ---
                # 3. MLP Expansion (c_fc)
                # 使用 mlp_init_scale
                elif name.endswith('.mlp.c_fc'):
                    target_std = config.mlp_init_scale
                
                # 4. MLP Output Projection (Residual Layer)
                # 使用 mlp_init_scale * resid_scale_factor
                elif name.endswith('.mlp.c_proj'):
                    target_std = config.mlp_init_scale * resid_scale_factor

                
                # --- Apply Initialization ---
                if target_std is not None:
                    
                    torch.nn.init.normal_(module.weight, mean=0.0, std=target_std)
                    if module.bias is not None:
                        torch.nn.init.zeros_(module.bias)
            
            # +++ NEW: Handle Embedding Layers +++
            elif isinstance(module, nn.Embedding):
                target_std = None
                # Check for wte or wpe (Token or Position embeddings)
                if name.endswith('.wte') or name.endswith('.wpe'):
                    target_std = config.emb_init_scale
                
                if target_std is not None:
                    torch.nn.init.normal_(module.weight, mean=0.0, std=target_std)
    
    
    # +++ MODIFIED: Replaced register_activation_hooks with a more general version +++
    def register_preconditioning_hooks(self):
        
        # --- Forward hook to capture input activations (X_t) ---
        def hook_fn_forward(module, input, output):
            x = input[0] 
            x = x.view(-1, x.size(-1))
            # Normalize by batch size for stability
            xxt = torch.matmul(x.t(), x) / x.shape[0]
            # Attach to the weight parameter
            module.weight.activation_second_moment = xxt

        # --- Backward hook to capture output gradients (G_Yt) ---
        def hook_fn_backward(module, grad_input, grad_output):
            gy = grad_output[0]
            gy = gy.view(-1, gy.size(-1))
            # Normalize by batch size for stability
            gygyt = torch.matmul(gy.t(), gy) / gy.shape[0]
            # Attach to the weight parameter
            module.weight.output_gradient_second_moment = gygyt

        print("Registering forward and backward hooks for preconditioning.")
        for name, module in self.named_modules():
            # Apply to all Linear layers except embeddings and the final head
            if isinstance(module, nn.Linear) and '.wte.' not in name and '.wpe.' not in name and 'lm_head' not in name:
                if hasattr(module, 'weight') and module.weight.dim() >= 2:
                    print(f"  - Hooks registered for: {name}")
                    module.register_forward_hook(hook_fn_forward)
                    module.register_full_backward_hook(hook_fn_backward)
    
    
    # +++ NEW: Method to register hooks for capturing activations for monitoring +++
    def register_monitoring_hooks(self):
        """
        Registers forward hooks on model layers to capture their input and output
        activations for detailed monitoring.
        """
        self.monitoring_activations.clear() # Clear previous activations

        def get_hook(name):
            def hook_fn(module, input, output):
                # Store both input and output activations
                if isinstance(input, tuple):
                    # For linear layers, input is a tuple
                    self.monitoring_activations[f"{name}.input"] = input[0].detach()
                else:
                    self.monitoring_activations[f"{name}.input"] = input.detach()
                
                if isinstance(output, tuple):
                    self.monitoring_activations[f"{name}.output"] = output[0].detach()
                else:
                    self.monitoring_activations[f"{name}.output"] = output.detach()
            return hook_fn

        print("Registering forward hooks for Activation Monitoring.")
        for name, module in self.named_modules():
            # We capture activations from Linear layers and LayerNorms
            if isinstance(module, (nn.Linear, LayerNorm)):
                 # We avoid the embedding layers and the final language model head
                # if '.wte.' not in name and '.wpe.' not in name and 'lm_head' not in name:
                #     print(f"  - Monitoring hook registered for: {name}")
                #     module.register_forward_hook(get_hook(name))
                # print(f"  - Monitoring hook registered for: {name}")
                module.register_forward_hook(get_hook(name))


    @contextlib.contextmanager
    def activation_logging_context(self):
        """
        在 Context 内注册 Hook，退出时自动移除。
        """
        hooks = []
        self.monitoring_activations = {} # 清空旧数据

        def get_hook(name):
            def hook_fn(module, input, output):
                # 你的 Hook 逻辑
                if isinstance(input, tuple):
                    self.monitoring_activations[f"{name}.input"] = input[0].detach()
                else:
                    self.monitoring_activations[f"{name}.input"] = input.detach()
                
                if isinstance(output, tuple):
                    self.monitoring_activations[f"{name}.output"] = output[0].detach()
                else:
                    self.monitoring_activations[f"{name}.output"] = output.detach()
            return hook_fn

        # 注册 Hooks
        for name, module in self.named_modules():
            # 这里可以保留你的过滤逻辑，例如只监控 Linear 层
            if isinstance(module, (nn.Linear, LayerNorm)): 
                 # 可以在这里加额外的过滤，例如排除 lm_head
                 hooks.append(module.register_forward_hook(get_hook(name)))
        
        try:
            yield # 执行被包裹的代码 (Forward Pass)
        finally:
            # 退出时移除所有 Hook
            for h in hooks:
                h.remove()

    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}"
        
        # +++ MODIFIED: Positional Embedding Logic +++
        tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
        
        if self.config.pos_emb_type == 'absolute':
            pos = torch.arange(0, t, dtype=torch.long, device=device) # shape (t)
            pos_emb = self.transformer.wpe(pos) # position embeddings of shape (t, n_embd)
            x = self.transformer.drop(tok_emb + pos_emb)
            # RoPE args are None
            freqs_cos, freqs_sin = None, None
            
        elif self.config.pos_emb_type == 'rope':
            # 不加 pos_emb，只做 dropout
            x = self.transformer.drop(tok_emb)
            # 准备 RoPE 需要的 freq table
            freqs_cos = self.freqs_cos
            freqs_sin = self.freqs_sin
        else:
            raise ValueError(f"Unknown pos_emb_type: {self.config.pos_emb_type}")
        # +++ END MODIFIED +++
        
               
        for block in self.transformer.h:
            x = block(x, freqs_cos=freqs_cos, freqs_sin=freqs_sin)
        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

    
    # +++ NEW: Helper methods for LoRA +++
    
    def freeze_all_parameters(self):
        """
        Freezes all parameters in the model *except* for LayerNorm parameters.
        This is called before applying LoRA adapters.
        """
        print("Freezing all parameters except LayerNorms...")
        for name, param in self.named_parameters():
            # 检查参数名是否属于 LayerNorm
            # 您的 LayerNorm 模块名为 'ln_1', 'ln_2', 'ln_f'
            if '.ln_' in name or '.ln_f.' in name:
                param.requires_grad = True
                print(f"  - Unfreezing (LayerNorm): {name}")
            else:
                param.requires_grad = False

    def apply_lora(self):
        """
        Recursively finds target nn.Linear layers and replaces them
        with LoRALinear layers.
        """
        config = self.config

        print(f"Applying LoRA (r={config.lora_r}, alpha={config.lora_alpha}) to target modules: {config.lora_target_modules}")
        
        target_modules = set(config.lora_target_modules)
        if not target_modules:
            print("Warning: `lora_target_modules` is empty. No LoRA layers will be applied.")
            return

        # Store (parent_module, child_name, old_module)
        modules_to_replace = []

        # First pass: find all modules to replace
        for name, module in self.named_modules():
            if isinstance(module, nn.Linear):
                # Check if the module's name ends with any target suffix
                for target_suffix in target_modules:
                    if name.endswith(target_suffix):
                        # Found a target
                        parent_name = ".".join(name.split('.')[:-1])
                        child_name = name.split('.')[-1]
                        
                        parent_module = self
                        if parent_name:
                            parent_module = dict(self.named_modules())[parent_name]
                            
                        modules_to_replace.append((parent_module, child_name, module))
                        break # Found a match, move to next module

        if not modules_to_replace:
            print("Warning: No matching nn.Linear layers found for target modules.")
            return

        # Second pass: perform the replacement
        for parent_module, child_name, old_module in modules_to_replace:
            # Create new LoRALinear layer
            new_lora_layer = LoRALinear(
                in_features=old_module.in_features,
                out_features=old_module.out_features,
                r=config.lora_r,
                lora_alpha=config.lora_alpha,
                lora_dropout=config.lora_dropout,
                bias=(old_module.bias is not None)
            )
            
            # Copy original weights and bias to the frozen .linear part
            new_lora_layer.linear.weight.data.copy_(old_module.weight.data)
            if old_module.bias is not None:
                new_lora_layer.linear.bias.data.copy_(old_module.bias.data)
                
            # Replace the old module with the new LoRA-enabled one
            setattr(parent_module, child_name, new_lora_layer)
            print(f"  - Replaced {name} with {new_lora_layer}")

    def print_trainable_parameters(self):
        """
        Prints a summary of trainable parameters in the model.
        """
        trainable_params = 0
        all_params = 0
        print("Trainable parameter summary:")
        for name, param in self.named_parameters():
            all_params += param.numel()
            if param.requires_grad:
                trainable_params += param.numel()
                print(f"  - {name} (shape: {param.shape}, numel: {param.numel()})")
        
        print(f"\nTotal trainable parameters: {trainable_params}")
        print(f"Total all parameters: {all_params}")
        print(f"Trainable percentage: {100 * trainable_params / all_params:.4f}%")
        
    def freeze_for_lora_resume(self):
        """
        Freezes all parameters *except* for LoRA parameters.
        Used when resuming a checkpoint that already has LoRA layers.
        """
        print("Freezing non-LoRA parameters for resume...")
        for name, param in self.named_parameters():
            if '.lora_A' not in name and '.lora_B' not in name:
                param.requires_grad = False
            else:
                # Ensure LoRA params are trainable (they should be already)
                param.requires_grad = True
                
    # +++ END NEW +++
    
    
    def configure_optimizers(self, config: OptimizerConfig, device_type: str, optimizer_name: str):
        """
        Configures the optimizer based on a strategy name and a configuration object.
        """
        # +++ MODIFIED: Import moved inside the method +++
        # This avoids import errors when this method is not called, e.g. during evaluation.
        from optimizer.optimizer_strategies import OPTIMIZER_STRATEGIES, OptimizerConfig

        # Determine if fused AdamW is available and update the config
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == 'cuda'
        config.extra_args = dict(fused=True) if use_fused else dict()

        print(f"Selected optimizer strategy: '{optimizer_name}'")
        if optimizer_name not in OPTIMIZER_STRATEGIES:
            raise ValueError(f"Unknown optimizer_name: '{optimizer_name}'. Available: {list(OPTIMIZER_STRATEGIES.keys())}")
        
        # Get the strategy function from the registry and execute it
        strategy_func = OPTIMIZER_STRATEGIES[optimizer_name]
        optimizer = strategy_func(self, config)

        print(f"Optimizer configured. Fused AdamW used: {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