Class: NanoGPT::Trainer

Inherits:
Object
  • Object
show all
Defined in:
lib/nano_gpt/trainer.rb

Overview

Training loop for GPT models

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(model:, data_loader:, config: {}) ⇒ Trainer

Returns a new instance of Trainer.



10
11
12
13
14
15
16
17
18
19
20
 
# File 'lib/nano_gpt/trainer.rb', line 10

def initialize(model:, data_loader:, config: {})
  @model = model
  @data_loader = data_loader
  @config = default_config.merge(config)

  @iter_num = 0
  @best_val_loss = Float::INFINITY

  setup_optimizer
  setup_lr_scheduler
end

Instance Attribute Details

#best_val_loss ⇒ Object (readonly)

Returns the value of attribute best_val_loss.



8
9
10
 
# File 'lib/nano_gpt/trainer.rb', line 8

def best_val_loss
  @best_val_loss
end

#config ⇒ Object (readonly)

Returns the value of attribute config.



8
9
10
 
# File 'lib/nano_gpt/trainer.rb', line 8

def config
  @config
end

#iter_num ⇒ Object (readonly)

Returns the value of attribute iter_num.



8
9
10
 
# File 'lib/nano_gpt/trainer.rb', line 8

def iter_num
  @iter_num
end

#model ⇒ Object (readonly)

Returns the value of attribute model.



8
9
10
 
# File 'lib/nano_gpt/trainer.rb', line 8

def model
  @model
end

#optimizer ⇒ Object (readonly)

Returns the value of attribute optimizer.



8
9
10
 
# File 'lib/nano_gpt/trainer.rb', line 8

def optimizer
  @optimizer
end

Instance Method Details

#default_config ⇒ Object 



22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
 
# File 'lib/nano_gpt/trainer.rb', line 22

def default_config
  {
    out_dir: "out",
    eval_interval: 250,
    log_interval: 10,
    eval_iters: 200,
    eval_only: false,
    always_save_checkpoint: false,

    # Optimizer
    learning_rate: 1e-3,
    weight_decay: 1e-1,
    beta1: 0.9,
    beta2: 0.99,
    grad_clip: 1.0,

    # LR scheduler
    decay_lr: true,
    warmup_iters: 100,
    lr_decay_iters: 5000,
    min_lr: 1e-4,

    # Training
    max_iters: 5000,
    gradient_accumulation_steps: 1,

    device: "cpu"
  }
end

#estimate_loss ⇒ Object 



114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
 
# File 'lib/nano_gpt/trainer.rb', line 114

def estimate_loss
  @model.eval
  out = {}

  [:train, :val].each do |split|
    losses = []
    @config[:eval_iters].times do
      x, y = @data_loader.get_batch(split)
      Torch.no_grad do
        _logits, loss = @model.call(x, targets: y)
        losses << loss.item
      end
    end
    out[split] = losses.sum / losses.size
  end

  @model.train
  out
end

#load_checkpoint(path) ⇒ Object 



153
154
155
156
157
158
159
160
161
162
163
164
165
 
# File 'lib/nano_gpt/trainer.rb', line 153

def load_checkpoint(path)
  checkpoint = Torch.load(path)

  @model.load_state_dict(checkpoint["model"])
  @iter_num = checkpoint["iter_num"]
  @best_val_loss = checkpoint["best_val_loss"]

  # Reinitialize optimizer (since we can't restore optimizer state in torch.rb)
  setup_optimizer

  puts "Loaded checkpoint from #{path} (iter #{@iter_num})"
  checkpoint
end

#save_checkpoint ⇒ Object 



134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
 
# File 'lib/nano_gpt/trainer.rb', line 134

def save_checkpoint
  FileUtils.mkdir_p(@config[:out_dir])
  path = File.join(@config[:out_dir], "ckpt.pt")

  # Note: torch.rb doesn't support optimizer.state_dict yet
  # We save model state and training metadata
  # Convert symbol keys to strings for Torch.save compatibility
  checkpoint = {
    "model" => @model.state_dict,
    "model_args" => stringify_keys(@model.config.to_h),
    "iter_num" => @iter_num,
    "best_val_loss" => @best_val_loss,
    "config" => stringify_keys(@config)
  }

  Torch.save(checkpoint, path)
  puts "Saved checkpoint to #{path}"
end

#train ⇒ Object 



52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
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
109
110
111
112
 
# File 'lib/nano_gpt/trainer.rb', line 52

def train
  puts "Starting training..."
  puts "Tokens per iteration: #{tokens_per_iter}"

  @model.train
  x, y = @data_loader.get_batch(:train)
  t0 = Time.now

  while @iter_num <= @config[:max_iters]
    # Set learning rate for this iteration
    lr = @config[:decay_lr] ? @lr_scheduler.step(@optimizer, @iter_num) : @config[:learning_rate]

    # Evaluate and checkpoint
    if @iter_num % @config[:eval_interval] == 0
      losses = estimate_loss
      puts "step #{@iter_num}: train loss #{losses[:train].round(4)}, val loss #{losses[:val].round(4)}"

      if losses[:val] < @best_val_loss || @config[:always_save_checkpoint]
        @best_val_loss = [losses[:val], @best_val_loss].min
        save_checkpoint if @iter_num > 0
      end
    end

    break if @iter_num == 0 && @config[:eval_only]

    # Forward/backward with gradient accumulation
    @optimizer.zero_grad

    accumulated_loss = 0.0
    @config[:gradient_accumulation_steps].times do |micro_step|
      logits, loss = @model.call(x, targets: y)
      loss = loss / @config[:gradient_accumulation_steps]
      accumulated_loss += loss.item
      loss.backward

      # Prefetch next batch
      x, y = @data_loader.get_batch(:train)
    end

    # Gradient clipping (manual implementation since torch.rb lacks clip_grad_norm_)
    if @config[:grad_clip] > 0.0
      clip_grad_norm(@model.parameters, @config[:grad_clip])
    end

    # Optimizer step
    @optimizer.step

    # Logging
    t1 = Time.now
    dt = t1 - t0
    t0 = t1

    if @iter_num % @config[:log_interval] == 0
      puts "iter #{@iter_num}: loss #{accumulated_loss.round(4)}, time #{(dt * 1000).round(2)}ms, lr #{lr.round(6)}"
    end

    @iter_num += 1
  end

  puts "Training complete!"
end