TensorRT-LLMs/tests/functional/test_layer_norm.py
2023-09-20 00:29:41 -07:00

106 lines
4.0 KiB
Python

# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
import unittest
import numpy as np
import pytest
import torch
from parameterized import parameterized
from polygraphy.backend.trt import CreateConfig, EngineFromNetwork, TrtRunner
import tensorrt_llm
from tensorrt_llm import Parameter, Tensor
from tensorrt_llm._utils import torch_to_numpy
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from utils.util import getSMVersion
class TestFunctional(unittest.TestCase):
def setUp(self):
tensorrt_llm.logger.set_level('error')
torch.manual_seed(42)
@parameterized.expand([['float16'], ['float32'], ['bfloat16']])
def test_layer_norm_plugin(self, dtype):
# Skip tests that are not supported in pre-ampere architecture
if getSMVersion() < 80:
if dtype == 'bfloat16':
pytest.skip(
"bfloat16 is not supported in pre-ampere architecture")
# test data
hidden_size = 1024
x_data = torch.randn((8, 128, hidden_size),
dtype=torch.float64,
device="cuda")
weight = torch.randn((hidden_size), dtype=torch.float64, device="cuda")
bias = torch.randn((hidden_size), dtype=torch.float64, device="cuda")
eps = 1e-5
m = torch.nn.LayerNorm(hidden_size,
eps=eps,
dtype=torch.float64,
device="cuda")
m.weight = torch.nn.Parameter(weight)
m.bias = torch.nn.Parameter(bias)
# pytorch run
with torch.no_grad():
ref = m(x_data)
m.to(tensorrt_llm._utils.str_dtype_to_torch(dtype))
x_data = x_data.to(tensorrt_llm._utils.str_dtype_to_torch(dtype))
gamma_data = m.weight.detach().cpu()
beta_data = m.bias.detach().cpu()
# construct trt network
builder = tensorrt_llm.Builder()
net = builder.create_network()
net.plugin_config.set_layernorm_plugin(dtype)
with tensorrt_llm.net_guard(net):
network = tensorrt_llm.default_trtnet()
x = Tensor(name='x',
shape=x_data.shape,
dtype=tensorrt_llm.str_dtype_to_trt(dtype))
weight = Parameter(torch_to_numpy(gamma_data.cpu())).value
bias = Parameter(torch_to_numpy(beta_data.cpu())).value
output = tensorrt_llm.functional.layer_norm(x, hidden_size, weight,
bias, eps).trt_tensor
output.name = 'output'
network.mark_output(output)
output.dtype = tensorrt_llm.str_dtype_to_trt(dtype)
# trt run
build_engine = EngineFromNetwork(
(builder.trt_builder, net.trt_network),
config=CreateConfig(fp16=(dtype == 'float16'),
bf16=(dtype == 'bfloat16')))
assert build_engine is not None, "Build engine failed"
with TrtRunner(build_engine) as runner:
outputs = runner.infer(feed_dict={'x': x_data.cpu()})
# compare diff
dtype_atol = {"float16": 2e-2, "float32": 2e-6, "bfloat16": 8e-2}
np.testing.assert_allclose(ref.cpu().numpy(),
outputs['output'].to(torch.float32),
atol=dtype_atol[dtype])