Module: GPT2ConfigLoader

Defined in:
lib/toy/io/gguf_load.rb

Class Method Summary collapse

Class Method Details

.read(path) ⇒ Object

Read all hyperparams from a GGUF file’s kv metadata. Returns a populated GPT2Config; the caller passes it to GPT2LM.new and the FFI cache realize_for methods.



41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
 
# File 'lib/toy/io/gguf_load.rb', line 41

def self.read(path)
  handle = TinyNN.tnn_gguf_load(path)
  if handle == nil
    puts "GPT2ConfigLoader: failed to open " + path
    return GPT2Config.new(0, 0, 0, 0, 0, 0, 1.0e-5)
  end
  vocab   = TinyNN.tnn_gguf_get_u32(handle, "gpt2.vocab_size")
  d_model = TinyNN.tnn_gguf_get_u32(handle, "gpt2.embedding_length")
  d_ff    = TinyNN.tnn_gguf_get_u32(handle, "gpt2.feed_forward_length")
  n_head  = TinyNN.tnn_gguf_get_u32(handle, "gpt2.attention.head_count")
  n_layer = TinyNN.tnn_gguf_get_u32(handle, "gpt2.block_count")
  ctx     = TinyNN.tnn_gguf_get_u32(handle, "gpt2.context_length")
  eps     = TinyNN.tnn_gguf_get_f32(handle, "gpt2.attention.layer_norm_epsilon")
  TinyNN.tnn_gguf_free(handle)
  GPT2Config.new(vocab, d_model, d_ff, n_head, n_layer, ctx, eps)
end