Class: Kiribi::Gemma4::E2B::Model

Inherits:
Object
  • Object
show all
Defined in:
lib/kiribi/gemma4/e2b/model.rb

Constant Summary collapse

EOS_TOKEN_IDS = 
[1, 106, 50]
IMAGE_TOKEN_ID = 
258_880
AUDIO_TOKEN_ID = 
258_881

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize ⇒ Model

Returns a new instance of Model.



20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 20

def initialize
  @tokenizer = Tokenizers.from_file(TOKENIZER_FILEPATH)
  @embed_model = OnnxRuntime::Model.new(EMBED_MODEL_FILEPATH)
  @decoder_model = OnnxRuntime::Model.new(DECODER_MODEL_FILEPATH)

  decoder_sess = OnnxRuntime::InferenceSession.new(DECODER_MODEL_FILEPATH)
  @head_dims = decoder_sess.inputs
    .select { it[:name].match?(/\Apast_key_values\.\d+\.key\z/) }
    .sort_by { it[:name][/\d+/].to_i }
    .map { it[:shape].last }
  @num_layers = @head_dims.length

  @num_logits_to_keep_1 = OnnxRuntime::OrtValue.from_shape_and_type([], :int64)
  @num_logits_to_keep_1.data_ptr.write_int64(1)
end

Instance Attribute Details

#tokenizer ⇒ Object (readonly)

Returns the value of attribute tokenizer.



18
19
20
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 18

def tokenizer
  @tokenizer
end

Instance Method Details

#chat(messages, max_new_tokens: 256) ⇒ Object 



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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 94

def chat(messages, max_new_tokens: 256)
  prompt_parts = ["<bos>"]
  encoded_media = []

  messages.each do |msg|
    role = msg[:role]
    content = msg[:content]
    prompt_parts << "<|turn>#{role}\n"

    if content.is_a?(String)
      prompt_parts << content
    elsif content.is_a?(Array)
      content.each do |part|
        case part[:type]
        when "text"
          prompt_parts << part[:text]
        when "image"
          features = part[:features]
          prompt_parts << "<|image>" + "<|image|>" * features.length + "<image|>\n"
          encoded_media << {token_id: IMAGE_TOKEN_ID, features:}
        when "audio"
          features = part[:features]
          prompt_parts << "<|audio>" + "<|audio|>" * features.length + "<audio|>\n"
          encoded_media << {token_id: AUDIO_TOKEN_ID, features:}
        end
      end
    end

    prompt_parts << "<turn|>\n"
  end
  prompt_parts << "<|turn>model\n"

  input_ids = tokenizer.encode(prompt_parts.join).ids

  embeds = []
  encoded_media.each do |media|
    positions = input_ids.each_with_index
      .select { |t, _| t == media[:token_id] }
      .map(&:last)
      .reject { |pos| embeds.any? { it[:pos] == pos } }
    media[:features].each_with_index do |feat, idx|
      break if idx >= positions.length
      embeds << {pos: positions[idx], feat:}
    end
  end

  past_kv = nil
  generated = []

  max_new_tokens.times do |step|
    cur_ids = step == 0 ? input_ids : [generated.last]
    seq_len = cur_ids.length
    total_len = input_ids.length + generated.length

    embed_out = embed(cur_ids)
    inputs_embeds = embed_out["inputs_embeds"]
    per_layer_inputs = embed_out["per_layer_inputs"]

    if step == 0
      embeds.each { inputs_embeds[0][it[:pos]] = it[:feat] }
    end

    result = forward(
      inputs_embeds:,
      per_layer_inputs:,
      attention_mask: [Array.new(total_len, 1)],
      position_ids: [(total_len - seq_len...total_len).to_a],
      past_key_values: past_kv,
    )
    past_kv = result[:past_key_values]

    next_token = result[:logits][0][-1].each_with_index.max_by { |v, _| v }[1]
    break if EOS_TOKEN_IDS.include?(next_token)
    generated << next_token
  end

  tokenizer.decode(generated)
end

#embed(input_ids) ⇒ Object

低レベル API: ONNX 呼び出しのみ



52
53
54
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 52

def embed(input_ids)
  @embed_model.predict({"input_ids" => [input_ids]})
end

#forward(inputs_embeds:, per_layer_inputs:, attention_mask:, position_ids:, past_key_values: nil) ⇒ Object 



56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 56

def forward(inputs_embeds:, per_layer_inputs:, attention_mask:, position_ids:, past_key_values: nil)
  past_kv = past_key_values || init_kv_cache
  input = {
    "inputs_embeds" => inputs_embeds,
    "attention_mask" => attention_mask,
    "position_ids" => position_ids,
    "num_logits_to_keep" => @num_logits_to_keep_1,
    "per_layer_inputs" => per_layer_inputs,
  }
  input.merge!(past_kv)
  out = @decoder_model.predict(input)

  new_kv = {}
  @num_layers.times do |i|
    new_kv["past_key_values.#{i}.key"] = out["present.#{i}.key"]
    new_kv["past_key_values.#{i}.value"] = out["present.#{i}.value"]
  end

  {logits: out["logits"], past_key_values: new_kv}
end

#generate(prompt, max_new_tokens: 256) ⇒ Object

高レベル API



90
91
92
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 90

def generate(prompt, max_new_tokens: 256)
  chat([{role: "user", content: prompt}], max_new_tokens:)
end

#init_kv_cache ⇒ Object 



77
78
79
80
81
82
83
84
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 77

def init_kv_cache
  kv = {}
  @num_layers.times do |i|
    kv["past_key_values.#{i}.key"] = OnnxRuntime::OrtValue.from_shape_and_type([1, 1, 0, @head_dims[i]], :float)
    kv["past_key_values.#{i}.value"] = OnnxRuntime::OrtValue.from_shape_and_type([1, 1, 0, @head_dims[i]], :float)
  end
  kv
end

#load_audio_encoder ⇒ Object 



44
45
46
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 44

def load_audio_encoder
  @audio_encoder ||= AudioEncoder.new
end

#load_vision_encoder ⇒ Object

遅延ロード（名前で初期化コストを明示）



40
41
42
 
# File 'lib/kiribi/gemma4/e2b/model.rb', line 40

def load_vision_encoder
  @vision_encoder ||= VisionEncoder.new
end