Class: Informers::PreTrainedModel

Inherits:
Object
  • Object
show all
Defined in:
lib/informers/models.rb

Constant Summary collapse

MAIN_INPUT_NAME = 
:input_ids

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(config, session) ⇒ PreTrainedModel

Returns a new instance of PreTrainedModel.



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/informers/models.rb', line 74

def initialize(config, session)
  super()

  @config = config
  @session = session

  @output_names = nil

  model_name = MODEL_CLASS_TO_NAME_MAPPING[self.class]
  model_type = MODEL_TYPE_MAPPING[model_name]

  case model_type
  when MODEL_TYPES[:DecoderOnly]
    @can_generate = true

    @run_beam = method(:decoder_run_beam)
    @get_start_beams = method(:decoder_start_beams)
    @update_beam = method(:decoder_update_beam)
    @forward = method(:decoder_forward)

  when MODEL_TYPES[:Seq2Seq], MODEL_TYPES[:Vision2Seq]
    @can_generate = true

    @run_beam = method(:seq2seq_run_beam)
    @get_start_beams = method(:seq2seq_start_beams)
    @update_beam = method(:seq2seq_update_beam)
    @forward = method(:seq2seq_forward)

  when MODEL_TYPES[:EncoderDecoder]
    @forward = method(:encoder_forward)

  else
    @forward = method(:encoder_forward)
  end
end

Instance Attribute Details

#config ⇒ Object (readonly)

Returns the value of attribute config.



72
73
74
 
# File 'lib/informers/models.rb', line 72

def config
  @config
end

Class Method Details

.construct_session(pretrained_model_name_or_path, file_name, **options) ⇒ Object 



178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
 
# File 'lib/informers/models.rb', line 178

def self.construct_session(pretrained_model_name_or_path, file_name, **options)
  prefix = "onnx/"
  if file_name.start_with?("../")
    prefix = ""
    file_name = file_name[3..]
  elsif file_name.start_with?("/")
    prefix = ""
    file_name = file_name[1..]
  end
  dtype = options[:dtype] || (options[:quantized] ? "q8" : "fp32")
  suffix = Utils::DEFAULT_DTYPE_SUFFIX_MAPPING[dtype.to_sym]
  if !suffix
    raise ArgumentError, "Invalid dtype: #{dtype}. Should be one of: #{Utils::DEFAULT_DTYPE_SUFFIX_MAPPING.keys.join(", ")}"
  end
  model_file_name = "#{prefix}#{file_name}#{suffix}.onnx"
  path = Utils::Hub.get_model_file(pretrained_model_name_or_path, model_file_name, true, **options)

  session_options = {
    providers: Backends::Onnx.device_to_execution_providers(options[:device]),
    log_severity_level: 4
  }.merge(options[:session_options] || {})

  begin
    OnnxRuntime::InferenceSession.new(path, **session_options)
  rescue OnnxRuntime::Error => e
    raise e unless e.message.include?("No such file or directory") && e.message.include?(".onnx_data")

    Utils::Hub.get_model_file(pretrained_model_name_or_path, "#{model_file_name}_data", true, **options)
    OnnxRuntime::InferenceSession.new(path, **session_options)
  end
end

.from_pretrained(pretrained_model_name_or_path, quantized: true, progress_callback: nil, config: nil, cache_dir: nil, local_files_only: false, revision: "main", device: nil, dtype: nil, model_file_name: nil, session_options: {}) ⇒ 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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
 
# File 'lib/informers/models.rb', line 110

def self.from_pretrained(
  pretrained_model_name_or_path,
  quantized: true,
  progress_callback: nil,
  config: nil,
  cache_dir: nil,
  local_files_only: false,
  revision: "main",
  device: nil,
  dtype: nil,
  model_file_name: nil,
  session_options: {}
)
  options = {
    quantized:,
    progress_callback:,
    config:,
    cache_dir:,
    local_files_only:,
    revision:,
    device:,
    dtype:,
    model_file_name:,
    session_options:
  }

  model_name = MODEL_CLASS_TO_NAME_MAPPING[self]
  model_type = MODEL_TYPE_MAPPING[model_name]

  config ||= AutoConfig.from_pretrained(pretrained_model_name_or_path, **options)

  if model_type == MODEL_TYPES[:DecoderOnly]
    info = [
      construct_session(pretrained_model_name_or_path, options[:model_file_name] || "decoder_model_merged", **options),
      Utils::Hub.get_model_json(pretrained_model_name_or_path, "generation_config.json", false, **options)
    ]

  elsif model_type == MODEL_TYPES[:Seq2Seq] || model_type == MODEL_TYPES[:Vision2Seq]
    info = [
      construct_session(pretrained_model_name_or_path, "encoder_model", **options),
      construct_session(pretrained_model_name_or_path, "decoder_model_merged", **options),
      Utils::Hub.get_model_json(pretrained_model_name_or_path, "generation_config.json", false, **options)
    ]

  elsif model_type == MODEL_TYPES[:MaskGeneration]
    info = [
      construct_session(pretrained_model_name_or_path, "vision_encoder", **options),
      construct_session(pretrained_model_name_or_path, "prompt_encoder_mask_decoder", **options)
    ]

  elsif model_type == MODEL_TYPES[:EncoderDecoder]
    info = [
      construct_session(pretrained_model_name_or_path, "encoder_model", **options),
      construct_session(pretrained_model_name_or_path, "decoder_model_merged", **options)
    ]

  else
    if model_type != MODEL_TYPES[:EncoderOnly]
      warn "Model type for '#{model_name || config[:model_type]}' not found, assuming encoder-only architecture. Please report this."
    end
    info = [
      construct_session(pretrained_model_name_or_path, options[:model_file_name] || "model", **options)
    ]
  end

  new(config, *info)
end

Instance Method Details

#call(model_inputs, **kwargs) ⇒ Object 



210
211
212
 
# File 'lib/informers/models.rb', line 210

def call(model_inputs, **kwargs)
  @forward.(model_inputs, **kwargs)
end

#generate(inputs, generation_config = nil, logits_processor = nil, inputs_attention_mask: nil) ⇒ Object 



214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
 
# File 'lib/informers/models.rb', line 214

def generate(inputs, generation_config = nil, logits_processor = nil, inputs_attention_mask: nil)
  if !@can_generate
    model_name = MODEL_CLASS_TO_NAME_MAPPING[self.class]
    error_message = "The current model class (#{model_name}) is not compatible with `.generate()`, as it doesn't have a language model head."
    raise Error, error_message
  end

  if !inputs.is_a?(Array)
    raise ArgumentError, "`inputs` must be an Array, but is #{inputs.class.name}"
  end

  if @config[:is_encoder_decoder]
    # Generating from the encoder outputs
    input_ids_seq_length = 0
  else
    input_ids_seq_length = inputs.length

    # decoder-only
    if input_ids_seq_length == 0
      raise Error, "Must supply a non-empty array of input token ids."
    end
  end

  # Update generation config with defaults
  generation_config = get_generation_config(generation_config)

  logits_processor ||= Utils::LogitsProcessorList.new

  # Update logits processor
  logits_processor = get_logits_processor(
    generation_config,
    input_ids_seq_length,
    logits_processor
  )

  eos_token_ids = generation_config[:eos_token_id]
  if !eos_token_ids.nil? && !eos_token_ids.is_a?(Array)
    eos_token_ids = [eos_token_ids]
  end

  num_output_tokens = 1
  max_output_tokens = num_output_tokens + (generation_config[:max_new_tokens] || Float::INFINITY)

  # Only use max length if max_new_tokens is not provided
  use_max_length = generation_config[:max_length].is_a?(Integer) && generation_config[:max_new_tokens].nil?
  sampler = Utils::Sampler.get_sampler(generation_config)

  beams = get_start_beams(inputs, generation_config, num_output_tokens, inputs_attention_mask)

  while beams.any? { |x| !x[:done] } && num_output_tokens < max_output_tokens
    newest_beams = []
    beams.each do |beam|
      if beam[:done]
        # Add this beam back into the pool
        newest_beams << beam
        next
      end
      if use_max_length && beam[:output_token_ids].length >= generation_config["max_length"]
        # Set this beam to done and add it back into the pool
        beam[:done] = true
        newest_beams << beam
        next
      end

      output = run_beam(beam)

      # add attentions/scores to beam only if user requested
      if generation_config["output_attentions"]
        add_attentions_to_beam(beam, output)
      end

      # Logits are of the form [batch_size, out_seq_length, vocab_size]
      # In most cases, this will be [batch_size, 1, vocab_size]
      # So, we select the last token's logits:
      # (equivalent to `logits = outputs.logits[:, -1, :]`)
      logits = output["logits"].map { |v| v[-1] }

      # Apply logits processor
      logits_processor.(beam[:output_token_ids], logits)

      sampled_tokens = sampler.(logits)
      sampled_tokens.each do |new_token_id, log_prob|
        # use previous beam as a starting point
        new_beam = beam.dup

        # update new beam
        update_beam(new_beam, new_token_id)

        new_beam[:score] += log_prob

        if eos_token_ids && eos_token_ids.include?(new_token_id)
          new_beam[:done] = true
        end

        newest_beams << new_beam
      end
    end
    num_output_tokens += 1

    # Next, we get the best beams, per ID
    newest_beams =
      group_beams(newest_beams).map do |group|
        group.sort_by { |v| -v[:score] }[0...generation_config["num_beams"]]
      end

    # Flatten beams
    beams = newest_beams.flatten(1)

    # Run callback
    if generation_config["callback_function"]
      generation_config["callback_function"].(beams)
    end
  end

  # TODO: Ensure that we can return non-batched outputs

  grouped_beams = group_beams(beams)

  get_flattened = lambda do |key|
    grouped_beams.flat_map do |batch|
      if generation_config["num_return_sequences"] > 1
        raise Todo
      else
        [batch[0][key]]
      end
    end
  end

  sequences = get_flattened.(:output_token_ids) # [1, seqLength]

  if generation_config["return_dict_in_generate"]
    raise Todo
  else
    sequences
  end
end