def test_seqlen_fp32(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp32'
for seqlen in [32, 128, 512]:
args_dict['max_seq_len'] = seqlen
max_diff = encoder_sample(args_dict)
tf.reset_default_graph()
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_batch_fp16(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp16'
for batch in [1, 8, 64, 128]:
args_dict['batch_size'] = batch
max_diff = encoder_sample(args_dict)
tf.reset_default_graph()
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_remove_padding_fp16(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp16'
for avg_seq_len in [32, 64, 128]:
args_dict['max_seq_len'] = 256
args_dict['remove_padding'] = 'True'
args_dict['avg_seq_len'] = avg_seq_len
max_diff = encoder_sample(args_dict)
tf.reset_default_graph()
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_batch_fp16(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp16'
for batch in [1, 8, 64, 128]:
args_dict['batch_size'] = batch
tf.reset_default_graph()
os.system("./bin/encoder_gemm {} {} {} {} {} 0".format(
args_dict['batch_size'], args_dict['max_seq_len'],
args_dict['head_number'], args_dict['size_per_head'],
args_dict['data_type'] == 'fp16'))
max_diff = encoder_sample(args_dict)
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_seqlen_fp32(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp32'
for seqlen in [32, 130, 512, 1024]:
args_dict['max_seq_len'] = seqlen
tf.reset_default_graph()
os.system("./bin/encoder_gemm {} {} {} {} {} 0".format(
args_dict['batch_size'], args_dict['max_seq_len'],
args_dict['head_number'], args_dict['size_per_head'],
args_dict['data_type'] == 'fp16'))
max_diff = encoder_sample(args_dict)
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_hidden_fp16(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp16'
for p in [
tuple([12, 64]),
tuple([16, 64]),
tuple([4, 32]),
tuple([8, 96])
]:
args_dict['head_number'] = p[0]
args_dict['size_per_head'] = p[1]
max_diff = encoder_sample(args_dict)
tf.reset_default_graph()
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])
def test_hidden_fp16(self):
args_dict = copy.deepcopy(self.common_args_dict)
args_dict['data_type'] = 'fp16'
for p in [
tuple([12, 64]),
tuple([16, 64]),
tuple([4, 32]),
tuple([8, 96])
]:
args_dict['head_number'] = p[0]
args_dict['size_per_head'] = p[1]
tf.reset_default_graph()
os.system("./bin/encoder_gemm {} {} {} {} {} 0".format(
args_dict['batch_size'], args_dict['max_seq_len'],
args_dict['head_number'], args_dict['size_per_head'],
args_dict['data_type'] == 'fp16'))
max_diff = encoder_sample(args_dict)
self.assertTrue(max_diff < self.threshold[args_dict['data_type']])