Class: Ignis::AI::Trainer

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

Overview

Trainer — complete training loop with gradient accumulation, checkpointing, and multi-GPU support via NvCCL.

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(model:, optimizer:, scheduler: nil, grad_accumulation_steps: 1, max_grad_norm: 1.0, use_nvccl: false, checkpoint_dir: nil) ⇒ Trainer

Returns a new instance of Trainer.

Parameters:

  • model (Transformer::Model)
  • optimizer (Optim::Base)
  • scheduler (Optim::LRScheduler::*, nil) (defaults to: nil)
  • grad_accumulation_steps (Integer) (defaults to: 1) —

    accumulate gradients over N steps

  • max_grad_norm (Float) (defaults to: 1.0) —

    gradient clipping norm

  • use_nvccl (Boolean) (defaults to: false) —

    enable multi-GPU gradient sync

  • checkpoint_dir (String, nil) (defaults to: nil) —

    directory for saving checkpoints



24
25
26
27
28
29
30
31
32
33
34
35
36
 
# File 'lib/nnw/ai/trainer.rb', line 24

def initialize(model:, optimizer:, scheduler: nil,
               grad_accumulation_steps: 1, max_grad_norm: 1.0,
               use_nvccl: false, checkpoint_dir: nil)
  @model = model
  @optimizer = optimizer
  @scheduler = scheduler
  @grad_accumulation_steps = grad_accumulation_steps
  @max_grad_norm = max_grad_norm
  @use_nvccl = use_nvccl
  @checkpoint_dir = checkpoint_dir
  @metrics = { steps: 0, total_loss: 0.0, best_loss: Float::INFINITY }
  @model.train!
end

Instance Attribute Details

#metrics ⇒ Hash (readonly)

Returns training metrics.

Returns:

  • (Hash) —

    training metrics



15
16
17
 
# File 'lib/nnw/ai/trainer.rb', line 15

def metrics
  @metrics
end

#model ⇒ Transformer::Model (readonly)

Returns:



9
10
11
 
# File 'lib/nnw/ai/trainer.rb', line 9

def model
  @model
end

#optimizer ⇒ Optim::Base (readonly)

Returns:



12
13
14
 
# File 'lib/nnw/ai/trainer.rb', line 12

def optimizer
  @optimizer
end

Instance Method Details

#load_checkpoint!(path) ⇒ void

This method returns an undefined value.

Load from checkpoint.

Parameters:

  • path (String) 


155
156
157
158
 
# File 'lib/nnw/ai/trainer.rb', line 155

def load_checkpoint!(path)
  Safetensors.load_model(@model, path, strict: false)
  Ignis.logger.info("Checkpoint loaded: #{path}")
end

#save_checkpoint! ⇒ String

Save model checkpoint.

Returns:

  • (String) —

    checkpoint path



131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
 
# File 'lib/nnw/ai/trainer.rb', line 131

def save_checkpoint!
  return unless @checkpoint_dir

  Dir.mkdir(@checkpoint_dir) unless Dir.exist?(@checkpoint_dir)
  path = File.join(@checkpoint_dir, "checkpoint_step_#{@metrics[:steps]}.safetensors")

  tensors = {}
  @model.named_parameters.each do |name, param|
    tensors[name] = param
  end

  Safetensors.save(tensors, path, metadata: {
    "step" => @metrics[:steps].to_s,
    "loss" => (@metrics[:total_loss] / [@metrics[:steps], 1].max).to_s,
    "framework" => "nnw"
  })

  Ignis.logger.info("Checkpoint saved: #{path}")
  path
end

#train(data_loader, steps:, log_interval: 100, checkpoint_interval: 1000, eval_fn: nil) ⇒ Hash

Train for a specified number of steps.

Parameters:

  • data_loader (DataLoader) —

    provides batches

  • steps (Integer) —

    total training steps

  • log_interval (Integer) (defaults to: 100) —

    log every N steps

  • checkpoint_interval (Integer) (defaults to: 1000) —

    save every N steps

  • eval_fn (Proc, nil) (defaults to: nil) —

    evaluation function called at log intervals

Returns:

  • (Hash) —

    final metrics



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
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
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
 
# File 'lib/nnw/ai/trainer.rb', line 45

def train(data_loader, steps:, log_interval: 100,
          checkpoint_interval: 1000, eval_fn: nil)
  @model.train!
  accumulated_loss = 0.0

  steps.times do |step|
    # Get batch
    batch = data_loader.next_batch
    input_ids = batch[:input_ids]
    targets = batch[:targets]

    # Forward pass
    logits = @model.call(input_ids)
    loss = Loss.cross_entropy(logits, targets)

    # Scale loss for gradient accumulation
    scaled_loss = loss * (1.0 / @grad_accumulation_steps)

    # Backward pass
    scaled_loss.backward!

    accumulated_loss += loss.item

    # Optimizer step (every grad_accumulation_steps)
    if (step + 1) % @grad_accumulation_steps == 0
      # Gradient clipping
      grad_norm = @optimizer.clip_grad_norm!(@max_grad_norm)

      # Multi-GPU gradient sync
      if @use_nvccl
        sync_gradients_nvccl!
      end

      # Optimizer step
      @optimizer.step
      @optimizer.zero_grad!
      @scheduler&.step

      @metrics[:steps] += 1
      @metrics[:total_loss] += accumulated_loss / @grad_accumulation_steps

      # Logging
      if @metrics[:steps] % log_interval == 0
        avg_loss = @metrics[:total_loss] / @metrics[:steps]
        lr = @optimizer.lr
        Ignis.logger.info(
          "Step #{@metrics[:steps]} | Loss: #{'%.4f' % (accumulated_loss / @grad_accumulation_steps)} | " \
          "Avg Loss: #{'%.4f' % avg_loss} | LR: #{'%.2e' % lr} | Grad Norm: #{'%.2f' % grad_norm}"
        )

        # EventBus publish
        if defined?(Ignis::Shared::EventBus)
          Ignis::Shared::EventBus.publish(:training_step, {
            step: @metrics[:steps],
            loss: accumulated_loss / @grad_accumulation_steps,
            avg_loss: avg_loss,
            lr: lr,
            grad_norm: grad_norm
          })
        end

        # Eval
        if eval_fn
          @model.eval!
          eval_fn.call(@model, @metrics[:steps])
          @model.train!
        end
      end

      # Checkpointing
      if @checkpoint_dir && @metrics[:steps] % checkpoint_interval == 0
        save_checkpoint!
      end

      accumulated_loss = 0.0
    end

    # Clear tape each iteration
    Tape.clear!
  end

  @metrics
end