Class: NanoGPT::GPT

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

Overview

GPT Language Model

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(config) ⇒ GPT

Returns a new instance of GPT.

Raises:

  • (ArgumentError) 


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
34
35
36
37
38
39
40
41
42
 
# File 'lib/nano_gpt/model.rb', line 8

def initialize(config)
  super()
  raise ArgumentError, "vocab_size must be set" unless config.vocab_size
  raise ArgumentError, "block_size must be set" unless config.block_size

  @config = config

  # Token and position embeddings
  @wte = Torch::NN::Embedding.new(config.vocab_size, config.n_embd)
  @wpe = Torch::NN::Embedding.new(config.block_size, config.n_embd)
  @drop = Torch::NN::Dropout.new(p: config.dropout)

  # Transformer blocks
  @h = Torch::NN::ModuleList.new(
    config.n_layer.times.map { Layers::Block.new(config) }
  )

  # Final layer norm
  @ln_f = Layers::LayerNorm.new(config.n_embd, bias: config.bias)

  # Note: We use weight tying - lm_head shares weights with wte
  # Instead of a separate Linear layer, we use wte.weight directly in forward

  # Initialize weights
  apply(method(:_init_weights))

  # Special scaled init for residual projections (per GPT-2 paper)
  named_parameters.each do |name, param|
    if name.end_with?("c_proj.weight")
      Torch::NN::Init.normal!(param, mean: 0.0, std: 0.02 / Math.sqrt(2 * config.n_layer))
    end
  end

  puts format("number of parameters: %.2fM", num_params / 1e6)
end

Instance Attribute Details

#config ⇒ Object (readonly)

Returns the value of attribute config.



6
7
8
 
# File 'lib/nano_gpt/model.rb', line 6

def config
  @config
end

Instance Method Details

#configure_optimizers(weight_decay:, learning_rate:, betas:, device_type:) ⇒ Object 



171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
 
# File 'lib/nano_gpt/model.rb', line 171

def configure_optimizers(weight_decay:, learning_rate:, betas:, device_type:)
  # Separate parameters into decay and no-decay groups
  # All 2D+ params (weights) get weight decay, 1D params (biases, layernorm) don't
  decay_params = []
  nodecay_params = []

  parameters.each do |param|
    next unless param.requires_grad

    if param.dim >= 2
      decay_params << param
    else
      nodecay_params << param
    end
  end

  num_decay = decay_params.sum(&:numel)
  num_nodecay = nodecay_params.sum(&:numel)
  puts "num decayed parameter tensors: #{decay_params.size}, with #{num_decay} parameters"
  puts "num non-decayed parameter tensors: #{nodecay_params.size}, with #{num_nodecay} parameters"

  # Create optimizer with parameter groups (using symbol keys for torch.rb)
  Torch::Optim::AdamW.new(
    [
      { params: decay_params, weight_decay: weight_decay },
      { params: nodecay_params, weight_decay: 0.0 }
    ],
    lr: learning_rate,
    betas: betas
  )
end

#crop_block_size(block_size) ⇒ Object

Raises:

  • (ArgumentError) 


149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
 
# File 'lib/nano_gpt/model.rb', line 149

def crop_block_size(block_size)
  raise ArgumentError, "Cannot crop to larger block_size" if block_size > @config.block_size

  @config.block_size = block_size

  # Create new embedding with cropped weights
  new_wpe = Torch::NN::Embedding.new(block_size, @config.n_embd)
  Torch.no_grad do
    new_wpe.weight.copy!(@wpe.weight[0...block_size])
  end
  @wpe = new_wpe

  # Update attention masks in all blocks
  @h.each do |block|
    attn = block.instance_variable_get(:@attn)
    next unless attn.instance_variable_defined?(:@mask)

    mask = attn.instance_variable_get(:@mask)
    attn.instance_variable_set(:@mask, mask[nil, nil, 0...block_size, 0...block_size])
  end
end

#estimate_mfu(fwdbwd_per_iter, dt) ⇒ Object

Estimate model flops utilization (MFU) See PaLM paper Appendix B: arxiv.org/abs/2204.02311



52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
 
# File 'lib/nano_gpt/model.rb', line 52

def estimate_mfu(fwdbwd_per_iter, dt)
  n = num_params
  cfg = @config
  l = cfg.n_layer
  h = cfg.n_head
  q = cfg.n_embd / cfg.n_head
  t = cfg.block_size

  # FLOPs per token and per forward-backward pass
  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 throughput as ratio of A100 bfloat16 peak FLOPS (312 TFLOPS)
  flops_achieved = flops_per_iter / dt
  flops_promised = 312e12
  flops_achieved / flops_promised
end

#forward(idx, targets: nil) ⇒ Object

Raises:

  • (ArgumentError) 


71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
 
# File 'lib/nano_gpt/model.rb', line 71

def forward(idx, targets: nil)
  b, t = idx.shape
  raise ArgumentError, "Sequence length #{t} exceeds block_size #{@config.block_size}" if t > @config.block_size

  device = idx.device

  # Position indices
  pos = Torch.arange(0, t, dtype: :long, device: device)

  # Embeddings
  tok_emb = @wte.call(idx)  # (B, T, n_embd)
  pos_emb = @wpe.call(pos)  # (T, n_embd)
  x = @drop.call(tok_emb + pos_emb)

  # Transformer blocks
  @h.each { |block| x = block.call(x) }

  # Final layer norm
  x = @ln_f.call(x)

  if targets
    # Training: compute logits for all positions using tied weights
    logits = Torch::NN::Functional.linear(x, @wte.weight, nil)
    loss = Torch::NN::Functional.cross_entropy(
      logits.view(-1, logits.size(-1)),
      targets.view(-1),
      ignore_index: -1
    )
  else
    # Inference: only compute logits for last position (optimization)
    # Use narrow to get last position: x[:, -1:, :]
    x_last = x.narrow(1, x.size(1) - 1, 1)
    logits = Torch::NN::Functional.linear(x_last, @wte.weight, nil)
    loss = nil
  end

  [logits, loss]
end

#generate(idx, max_new_tokens, temperature: 1.0, top_k: nil) ⇒ Object 



110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
 
# File 'lib/nano_gpt/model.rb', line 110

def generate(idx, max_new_tokens, temperature: 1.0, top_k: nil)
  Torch.no_grad do
    max_new_tokens.times do
      # Crop context if exceeds block_size
      idx_cond = if idx.size(1) <= @config.block_size
                   idx
                 else
                   idx.narrow(1, idx.size(1) - @config.block_size, @config.block_size)
                 end

      # Forward pass
      logits, _loss = forward(idx_cond)

      # Get logits for last position and scale by temperature
      # logits shape is (B, 1, vocab_size), squeeze to (B, vocab_size)
      logits = logits.squeeze(1) / temperature

      # Optional top-k filtering
      if top_k
        k = [top_k, logits.size(-1)].min
        v, _indices = logits.topk(k)
        # Get the k-th largest value as threshold
        threshold = v.narrow(1, k - 1, 1)
        # Mask out values below threshold
        logits = logits.masked_fill(logits.lt(threshold), -Float::INFINITY)
      end

      # Sample from probability distribution
      probs = Torch::NN::Functional.softmax(logits, dim: -1)
      idx_next = Torch.multinomial(probs, num_samples: 1)

      # Append to sequence
      idx = Torch.cat([idx, idx_next], dim: 1)
    end
  end

  idx
end

#num_params(non_embedding: true) ⇒ Object 



44
45
46
47
48
 
# File 'lib/nano_gpt/model.rb', line 44

def num_params(non_embedding: true)
  n_params = parameters.sum(&:numel)
  n_params -= @wpe.weight.numel if non_embedding
  n_params
end