modeling_snowflake.py

import torch
import torch.nn as nn
import math

# --- Submodule: FusedQKVAttention ---
class FusedQKVAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads
        # Fused QKV projection
        self.qkv_proj = nn.Linear(d_model, 3 * d_model)
        self.wo = nn.Linear(d_model, d_model)
        # Initialize weights for better training stability
        nn.init.xavier_uniform_(self.qkv_proj.weight)
        nn.init.xavier_uniform_(self.wo.weight)
        nn.init.zeros_(self.qkv_proj.bias)
        nn.init.zeros_(self.wo.bias)

    def forward(self, x, attention_mask=None):
        batch_size, seq_len, _ = x.shape
        # Fused projection and reshape
        qkv = self.qkv_proj(x).reshape(batch_size, seq_len, 3, self.num_heads, self.head_dim)
        qkv = qkv.permute(2, 0, 3, 1, 4)  # [3, batch, heads, seq_len, head_dim]
        q, k, v = qkv[0], qkv[1], qkv[2]
        # Compute attention with memory efficiency
        attention_scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        if attention_mask is not None:
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
            attention_scores = attention_scores.masked_fill(attention_mask == 0, float('-inf'))
        attention_weights = torch.softmax(attention_scores, dim=-1)
        # Apply attention and reshape
        context = torch.matmul(attention_weights, v)
        context = context.transpose(1, 2).reshape(batch_size, seq_len, self.d_model)
        return self.wo(context)


# --- Submodule: EnhancedFeedForward ---
class EnhancedFeedForward(nn.Module):
    def __init__(self, d_model, ff_dim, dropout=0.1):
        super().__init__()
        self.linear1 = nn.Linear(d_model, ff_dim)
        self.dropout1 = nn.Dropout(dropout)
        self.linear2 = nn.Linear(ff_dim, d_model)
        self.dropout2 = nn.Dropout(dropout)
        self.activation = nn.GELU()
        # Initialize weights for better training
        nn.init.xavier_uniform_(self.linear1.weight)
        nn.init.xavier_uniform_(self.linear2.weight)
        nn.init.zeros_(self.linear1.bias)
        nn.init.zeros_(self.linear2.bias)

    def forward(self, x):
        return self.dropout2(self.linear2(self.dropout1(self.activation(self.linear1(x)))))


# --- Submodule: EnhancedTransformerBlock ---
class EnhancedTransformerBlock(nn.Module):
    def __init__(self, d_model, num_heads, ff_dim, dropout=0.1):
        super().__init__()
        self.attention = FusedQKVAttention(d_model, num_heads)
        self.norm1 = nn.LayerNorm(d_model, eps=1e-6)
        self.dropout1 = nn.Dropout(dropout)
        self.feed_forward = EnhancedFeedForward(d_model, ff_dim, dropout)
        self.norm2 = nn.LayerNorm(d_model, eps=1e-6)
        self.dropout2 = nn.Dropout(dropout)

    def forward(self, x, attention_mask=None):
        # Pre-norm architecture
        attn_input = self.norm1(x)
        attn_output = self.attention(attn_input, attention_mask)
        x = x + self.dropout1(attn_output)
        ff_input = self.norm2(x)
        ff_output = self.feed_forward(ff_input)
        x = x + self.dropout2(ff_output)
        return x


# --- Main Model Class: Snowflake4CausalLM ---
class Snowflake4CausalLM(nn.Module):
    def __init__(self, vocab_size, max_seq_length, d_model, num_heads, num_layers, ff_dim, dropout=0.1):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, d_model)
        # Initialize positional encodings without in-place modification
        self.pos_encoding = nn.Parameter(torch.zeros(1, max_seq_length, d_model))
        position = torch.arange(max_seq_length).unsqueeze(1).float()
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pos_enc = torch.zeros(1, max_seq_length, d_model)
        pos_enc[0, :, 0::2] = torch.sin(position * div_term)
        pos_enc[0, :, 1::2] = torch.cos(position * div_term)
        self.pos_encoding.data = pos_enc.data
        self.layers = nn.ModuleList([
            EnhancedTransformerBlock(d_model, num_heads, ff_dim, dropout)
            for _ in range(num_layers)
        ])
        self.final_norm = nn.LayerNorm(d_model, eps=1e-6)
        self.dropout = nn.Dropout(dropout)
        self.fc_out = nn.Linear(d_model, vocab_size)
        # Tie embedding and output weights for memory efficiency and better generalization
        self.fc_out.weight = self.embedding.weight
        # Initialize embedding weights
        nn.init.normal_(self.embedding.weight, mean=0, std=0.02)

    def forward(self, input_ids, attention_mask=None):
        seq_length = input_ids.size(1)
        x = self.embedding(input_ids) + self.pos_encoding[:, :seq_length, :]
        x = self.dropout(x)
        for layer in self.layers:
            x = layer(x, attention_mask)
        x = self.final_norm(x)
        return self.fc_out(x)