MCPcopy Create free account
hub / github.com/THUDM/GLM / __init__

Method __init__

mpu/transformer.py:362–378  ·  view source on GitHub ↗
(self, hidden_size, output_dropout_prob, init_method,
                 output_layer_init_method=None)

Source from the content-addressed store, hash-verified

360 """
361
362 def __init__(self, hidden_size, output_dropout_prob, init_method,
363 output_layer_init_method=None):
364 super(ParallelMLP, self).__init__()
365 # Set output layer initialization if not provided.
366 if output_layer_init_method is None:
367 output_layer_init_method = init_method
368 # Project to 4h.
369 self.dense_h_to_4h = ColumnParallelLinear(hidden_size, 4 * hidden_size,
370 gather_output=False,
371 init_method=init_method)
372 # Project back to h.
373 self.dense_4h_to_h = RowParallelLinear(
374 4 * hidden_size,
375 hidden_size,
376 input_is_parallel=True,
377 init_method=output_layer_init_method)
378 self.dropout = torch.nn.Dropout(output_dropout_prob)
379
380 def forward(self, hidden_states):
381 # [b, s, 4hp]

Callers 6

__init__Method · 0.45
__init__Method · 0.45
__init__Method · 0.45
__init__Method · 0.45
__init__Method · 0.45
__init__Method · 0.45

Calls 2

RowParallelLinearClass · 0.85

Tested by

no test coverage detected