Files
sheepOp/models/attention.py
Carlos Gutierrez 3d2da94ce2 Initial commit: SheepOp LLM - Transformer-based language model implementation
- Complete transformer implementation from scratch
- Training pipeline with gradient accumulation and mixed precision
- Optimized inference with KV caching
- Multi-format data processing (PDFs, images, code, text)
- Comprehensive documentation
- Apache 2.0 license
- Example training plots included in docs/images/
2025-11-06 22:07:41 -05:00

221 lines
7.2 KiB
Python

"""
Multi-Head Attention mechanism from "Attention Is All You Need"
Includes optimizations for long context and hallucination reduction
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Tuple
class MultiHeadAttention(nn.Module):
"""
Multi-Head Attention mechanism with optional causal masking.
Features:
- Scaled dot-product attention
- Optional causal masking for autoregressive generation
- Efficient attention computation
"""
def __init__(
self,
d_model: int,
num_heads: int,
dropout: float = 0.1,
bias: bool = False,
causal: bool = False,
):
"""
Args:
d_model: Model dimension
num_heads: Number of attention heads
dropout: Dropout probability
bias: Whether to use bias in linear layers
causal: Whether to use causal masking
"""
super().__init__()
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
self.d_model = d_model
self.num_heads = num_heads
self.d_k = d_model // num_heads
self.causal = causal
# Linear projections for Q, K, V
self.q_proj = nn.Linear(d_model, d_model, bias=bias)
self.k_proj = nn.Linear(d_model, d_model, bias=bias)
self.v_proj = nn.Linear(d_model, d_model, bias=bias)
self.out_proj = nn.Linear(d_model, d_model, bias=bias)
self.dropout = nn.Dropout(dropout)
self.scale = 1.0 / math.sqrt(self.d_k)
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass of multi-head attention.
Args:
query: Query tensor [batch_size, seq_len, d_model]
key: Key tensor [batch_size, seq_len, d_model]
value: Value tensor [batch_size, seq_len, d_model]
mask: Optional attention mask [batch_size, seq_len, seq_len]
Returns:
output: Attention output [batch_size, seq_len, d_model]
attention_weights: Attention weights [batch_size, num_heads, seq_len, seq_len]
"""
batch_size, seq_len, _ = query.shape
# Project Q, K, V
Q = self.q_proj(query) # [batch_size, seq_len, d_model]
K = self.k_proj(key)
V = self.v_proj(value)
# Reshape for multi-head attention
Q = Q.view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2) # [batch_size, num_heads, seq_len, d_k]
K = K.view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2)
V = V.view(batch_size, seq_len, self.num_heads, self.d_k).transpose(1, 2)
# Compute attention scores
scores = torch.matmul(Q, K.transpose(-2, -1)) * self.scale # [batch_size, num_heads, seq_len, seq_len]
# Apply causal mask if needed
if self.causal:
causal_mask = torch.triu(
torch.ones(seq_len, seq_len, device=query.device, dtype=torch.bool),
diagonal=1
)
scores.masked_fill_(causal_mask, float('-inf'))
# Apply external mask if provided
if mask is not None:
scores.masked_fill_(mask.unsqueeze(1) == 0, float('-inf'))
# Compute attention weights
attention_weights = F.softmax(scores, dim=-1)
attention_weights = self.dropout(attention_weights)
# Apply attention to values
output = torch.matmul(attention_weights, V) # [batch_size, num_heads, seq_len, d_k]
# Concatenate heads
output = output.transpose(1, 2).contiguous() # [batch_size, seq_len, num_heads, d_k]
output = output.view(batch_size, seq_len, self.d_model) # [batch_size, seq_len, d_model]
# Final projection
output = self.out_proj(output)
return output, attention_weights
class PositionalEncoding(nn.Module):
"""
Positional encoding for transformer models.
Uses sinusoidal positional encoding as described in "Attention Is All You Need".
"""
def __init__(self, d_model: int, max_len: int = 5000, dropout: float = 0.1):
"""
Args:
d_model: Model dimension
max_len: Maximum sequence length
dropout: Dropout probability
"""
super().__init__()
self.dropout = nn.Dropout(p=dropout)
# Create positional encoding matrix
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0) # [1, max_len, d_model]
# Register as buffer (not a parameter)
self.register_buffer('pe', pe)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Add positional encoding to input.
Args:
x: Input tensor [batch_size, seq_len, d_model]
Returns:
Output with positional encoding added
"""
seq_len = x.shape[1]
x = x + self.pe[:, :seq_len, :]
return self.dropout(x)
class RotaryPositionalEncoding(nn.Module):
"""
Rotary Position Embedding (RoPE) - More efficient for long sequences.
Better for long-horizon execution tasks.
"""
def __init__(self, d_model: int, max_len: int = 8192):
"""
Args:
d_model: Model dimension (must be even)
max_len: Maximum sequence length
"""
super().__init__()
assert d_model % 2 == 0, "d_model must be even for RoPE"
self.d_model = d_model
self.max_len = max_len
# Precompute frequency matrix
inv_freq = 1.0 / (10000 ** (torch.arange(0, d_model, 2).float() / d_model))
self.register_buffer('inv_freq', inv_freq)
def forward(self, x: torch.Tensor, offset: int = 0) -> torch.Tensor:
"""
Apply rotary positional encoding.
Args:
x: Input tensor [batch_size, seq_len, d_model]
offset: Position offset for relative positions
Returns:
Rotated input tensor
"""
seq_len = x.shape[1]
device = x.device
# Generate position indices
t = torch.arange(offset, offset + seq_len, device=device).type_as(self.inv_freq)
freqs = torch.outer(t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
# Apply rotation
cos = emb.cos()
sin = emb.sin()
# Split input into two halves
x1, x2 = x.chunk(2, dim=-1)
# Apply rotation
rotated = torch.cat([
x1 * cos - x2 * sin,
x1 * sin + x2 * cos
], dim=-1)
return rotated