Class: NanoGPT::Layers::CausalSelfAttention

Inherits:
Torch::NN::Module
  • Object
show all
Defined in:
lib/nano_gpt/layers/causal_self_attention.rb

Overview

Multi-head causal self-attention

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ CausalSelfAttention

Returns a new instance of CausalSelfAttention.

Raises:

  • (ArgumentError) 


7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
 
# File 'lib/nano_gpt/layers/causal_self_attention.rb', line 7

def initialize(config)
  super()
  raise ArgumentError, "n_embd must be divisible by n_head" unless (config.n_embd % config.n_head).zero?

  @n_head = config.n_head
  @n_embd = config.n_embd
  @head_size = config.n_embd / config.n_head
  @dropout_p = config.dropout

  # Key, query, value projections for all heads, combined
  @c_attn = Torch::NN::Linear.new(config.n_embd, 3 * config.n_embd, bias: config.bias)
  # Output projection
  @c_proj = Torch::NN::Linear.new(config.n_embd, config.n_embd, bias: config.bias)
  # Regularization
  @attn_dropout = Torch::NN::Dropout.new(p: config.dropout)
  @resid_dropout = Torch::NN::Dropout.new(p: config.dropout)

  # Use native scaled_dot_product_attention with is_causal=true when dropout is 0
  # Native SDPA is ~5x faster but doesn't support dropout with is_causal mode
  @flash = config.dropout == 0.0

  # Causal mask for manual attention (only used when @flash is false)
  unless @flash
    mask = Torch.tril(Torch.ones(config.block_size, config.block_size))
    register_buffer("mask", mask.view(1, 1, config.block_size, config.block_size))
  end
end

Instance Method Details

#forward(x) ⇒ Object 



35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
 
# File 'lib/nano_gpt/layers/causal_self_attention.rb', line 35

def forward(x)
  b, t, c = x.shape

  # Calculate Q, K, V
  qkv = @c_attn.call(x)
  q, k, v = qkv.split(@n_embd, 2)

  # Reshape: (B, T, C) -> (B, nh, T, hs)
  q = q.view(b, t, @n_head, @head_size).transpose(1, 2)
  k = k.view(b, t, @n_head, @head_size).transpose(1, 2)
  v = v.view(b, t, @n_head, @head_size).transpose(1, 2)

  y = if @flash
        # Native scaled_dot_product_attention with is_causal=true
        # Uses Flash Attention on CUDA, optimized kernel on MPS
        Torch::NN.scaled_dot_product_attention(q, k, v, nil, 0.0, true)
      else
        # Manual attention implementation with causal mask
        scale = 1.0 / Math.sqrt(@head_size)
        att = q.matmul(k.transpose(-2, -1))
        att.mul!(scale)

        # Apply causal mask - slice mask to current sequence length
        mask_slice = @mask.narrow(2, 0, t).narrow(3, 0, t)
        att.masked_fill!(mask_slice.eq(0), -Float::INFINITY)
        att = Torch::NN::Functional.softmax(att, dim: -1)
        att = @attn_dropout.call(att)
        att.matmul(v)
      end

  # Reassemble heads: (B, nh, T, hs) -> (B, T, C)
  y = y.transpose(1, 2).contiguous.view(b, t, c)

  # Output projection
  @resid_dropout.call(@c_proj.call(y))
end