ag4masses/alphageometry/geometry_150M_generate.gin

NUM_EMBEDDINGS = 1024

# Number of parameters = 152M
NUM_LAYERS = 12
EMBED_DIM = 1024
NUM_HEADS = 8
HEAD_DIM = 128
MLP_DIM = 4096


transformer_layer.TransformerLayerGenerate:
  num_heads = %NUM_HEADS
  head_size = %HEAD_DIM
  window_length = 1024
  use_long_xl_architecture = False
  max_unrolled_windows = -1           # Always unroll.
  relative_position_type = "t5"       # Can be "fourier", "t5", or None.
  use_causal_mask = True
  attn_dropout_rate = %ATTN_DROPOUT_RATE   # Attention matrix dropout.
  memory_num_neighbors = 0
  dtype = %DTYPE

decoder_stack.DecoderStackGenerate:
  num_layers = %NUM_LAYERS
  embedding_size = %EMBED_DIM
  embedding_stddev = 1.0
  layer_factory = @transformer_layer.TransformerLayerGenerate
  dstack_window_length = 0
  use_absolute_positions = False
  use_final_layernorm = True          # Final layernorm before token lookup.
  final_dropout_rate = %DROPOUT_RATE  # Dropout before token lookup.
  final_mlp_factory = None            # Final MLP to predict target tokens.
  recurrent_layer_indices = ()
  memory_factory = None     # e.g. @memory_factory.memory_on_tpu_factory
  memory_layer_indices = ()
  dtype = %DTYPE


models.DecoderOnlyLanguageModelGenerate:
  num_heads = %NUM_HEADS
  head_size = %HEAD_DIM
  task_config = @decoder_stack.TransformerTaskConfig()
  decoder_factory = @decoder_stack.DecoderStackGenerate


training_loop.Trainer:
  model_definition = @models.DecoderOnlyLanguageModelGenerate