def test_create_attention_mask(self): config = GPTNeoConfig.from_pretrained("valhalla/gpt-neo-random-tiny") window_size = config.window_size batch_size, seq_length = 8, 1 block_length, num_blocks = GPTNeoAttentionMixin._get_block_length_and_num_blocks( seq_length, window_size) # causal_mask = layer._create_attention_mask(batch_size, seq_length, num_blocks, block_length, torch_device) causal_mask = GPTNeoAttentionMixin.create_local_attention_mask( batch_size, seq_length, config.window_size, torch_device) # check shapes expected_shape = [ batch_size, num_blocks, 1, block_length, window_size + block_length ] self.assertListEqual(list(causal_mask.shape), expected_shape) # first window_size tokens in the first block are always padded # and should not be attended self.assertTrue(torch.all(causal_mask[:, 0, :, :, :window_size] == 0)) # each window can attend at most window_size tokens self.assertTrue( torch.all(torch.sum(causal_mask, dim=4) <= config.window_size)) # check if user provided attention_mask is handled correctly attention_mask = torch.ones(batch_size, seq_length, dtype=torch.long, device=torch_device) attention_mask[:, -3:] = 0 # don't attend last 3 tokens # causal_mask = layer._create_attention_mask( # batch_size, seq_length, num_blocks, block_length, torch_device, attention_mask # ) causal_mask = GPTNeoAttentionMixin.create_local_attention_mask( batch_size, seq_length, config.window_size, torch_device, attention_mask) # last 3 tokens will be in the last block and shoul have 0s in causal_mask self.assertTrue(torch.all(causal_mask[:, -1, :, :, -3:] == 0)) # check shapes expected_shape = [ batch_size, num_blocks, 1, block_length, window_size + block_length ] self.assertListEqual(list(causal_mask.shape), expected_shape) # first window_size tokens in the first block are always padded # and should not be attended self.assertTrue(torch.all(causal_mask[:, 0, :, :, :window_size] == 0)) # each window can attend at most window_size tokens self.assertTrue( torch.all(torch.sum(causal_mask, dim=4) <= config.window_size))
def test_local_attn_probs(self): model = GPTNeoModel.from_pretrained( "valhalla/gpt-neo-random-tiny").eval() layer = model.h[1].attn.attention.to(torch_device) hidden_states = self._get_hidden_states() hidden_states = torch.cat([hidden_states, hidden_states - 0.5], dim=2) batch_size, seq_length, hidden_size = hidden_states.shape mask_tokens = 3 attention_mask = torch.ones(batch_size, seq_length, device=torch_device, dtype=torch.long) attention_mask[:, -mask_tokens:] = 0 # dont atten last mask_tokens local_causal_mask = GPTNeoAttentionMixin.create_local_attention_mask( batch_size, seq_length, model.config.window_size, torch_device, attention_mask) _, attn_probs = layer(hidden_states, attention_mask=local_causal_mask, output_attentions=True) # the last 3 tokens will be in the last block, and should have 0 attn_probs self.assertTrue( torch.all(attn_probs[:, -1, :, -mask_tokens:, -mask_tokens:] == 0)) # the first config.window_size tokens in the first block are always padded # and should have 0 attn_probs self.assertTrue( torch.all(attn_probs[:, 0, :, :model.config.window_size:, :model. config.window_size] == 0))
def test_look_back(self): hidden_states = self._get_hidden_states() batch_size, seq_length, hidden_size = hidden_states.shape # check when seq_length is divisible by window_size window_size = 4 block_length, num_block = GPTNeoAttentionMixin._get_block_length_and_num_blocks( seq_length, window_size) blocked_hidden_states = GPTNeoAttentionMixin._look_back( hidden_states, block_length, window_size) expected_shape = [ batch_size, num_block, window_size + block_length, hidden_size ] self.assertListEqual(list(blocked_hidden_states.shape), expected_shape) # The last block should contain the last (window_size + block_length) hidden_states self.assertTrue( torch.all(blocked_hidden_states[:, -1, ...] == hidden_states[:, -(window_size + block_length):, ...])) # check when seq_length is not divisible by window_size window_size = 3 block_length, num_block = GPTNeoAttentionMixin._get_block_length_and_num_blocks( seq_length, window_size) blocked_hidden_states = GPTNeoAttentionMixin._look_back( hidden_states, block_length, window_size) expected_shape = [ batch_size, num_block, window_size + block_length, hidden_size ] self.assertListEqual(list(blocked_hidden_states.shape), expected_shape) # The last block should contain the last (window_size + block_length) hidden_states self.assertTrue( torch.all(blocked_hidden_states[:, -1, ...] == hidden_states[:, -(window_size + block_length):, ...])) # check when window_size is > seq_length window_size = 19 block_length, num_block = GPTNeoAttentionMixin._get_block_length_and_num_blocks( seq_length, window_size) blocked_hidden_states = GPTNeoAttentionMixin._look_back( hidden_states, block_length, window_size) expected_shape = [ batch_size, num_block, window_size + block_length, hidden_size ] self.assertListEqual(list(blocked_hidden_states.shape), expected_shape) # when window_size > seq_length, num_blocks becomes 1, in this case # the first window_size values in blocked_hidden_staes are all zeros # and the last block_length values are equal to the hidden_states values = blocked_hidden_states[:, -1, :window_size, ...] expected_values = torch.zeros_like(values) self.assertTrue(torch.all(values == expected_values)) self.assertTrue( torch.all(blocked_hidden_states[:, -1, -block_length:, ...] == hidden_states))