mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
123 lines
4.4 KiB
Python
123 lines
4.4 KiB
Python
# SPDX-FileCopyrightText: Copyright (c) 2022-2024 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 unittest
|
|
|
|
import numpy as np
|
|
|
|
# isort: off
|
|
import torch
|
|
# isort: on
|
|
from polygraphy.backend.trt import EngineFromNetwork, TrtRunner
|
|
|
|
import tensorrt_llm
|
|
from tensorrt_llm import Tensor
|
|
|
|
|
|
class TestFunctional(unittest.TestCase):
|
|
|
|
def setUp(self):
|
|
tensorrt_llm.logger.set_level('warning')
|
|
|
|
def ref_categorical_sample(self, probs: torch.Tensor):
|
|
probs = probs / probs.sum(-1, keepdim=True)
|
|
rand_data = torch.rand(probs.shape[0],
|
|
dtype=probs.dtype,
|
|
device=probs.device)
|
|
cum_probs = probs.cumsum(-1)
|
|
samples = (cum_probs >= rand_data.unsqueeze(1)).int().argmax(dim=-1)
|
|
# print(samples)
|
|
return samples
|
|
|
|
# @unittest.skip("")
|
|
def test_ref_sample(self):
|
|
bs = 2
|
|
nbins = 10
|
|
probs = torch.rand((bs, nbins), dtype=torch.float32)
|
|
scaled_probs = probs / probs.sum(-1, keepdim=True)
|
|
print(scaled_probs)
|
|
samples = []
|
|
reps = 20000
|
|
for _ in range(reps):
|
|
samples.append(self.ref_categorical_sample(probs))
|
|
samples = torch.stack(samples).float()
|
|
# print(samples[:, 0], samples[:, 1])
|
|
hist = []
|
|
bins = torch.arange(nbins + 1).float()
|
|
for i in range(bs):
|
|
h = torch.histogram(samples[:, i], bins=bins).hist
|
|
h = h / h.sum(-1)
|
|
hist.append(h)
|
|
np.testing.assert_allclose(torch.stack(hist), scaled_probs, atol=1e-2)
|
|
return
|
|
|
|
def test_sample(self):
|
|
# test data
|
|
bs = 2
|
|
nbins = 10
|
|
probs = torch.rand((bs, nbins), dtype=torch.float32)
|
|
scaled_probs = probs / probs.sum(-1, keepdim=True)
|
|
print(scaled_probs)
|
|
|
|
# construct trt network
|
|
builder = tensorrt_llm.Builder()
|
|
net = builder.create_network()
|
|
with tensorrt_llm.net_guard(net):
|
|
network = tensorrt_llm.default_trtnet()
|
|
x = Tensor(name='x',
|
|
shape=probs.shape,
|
|
dtype=tensorrt_llm.torch_dtype_to_trt(probs.dtype))
|
|
# NOTE: we need rand() here since TRT rand() produces same numbers
|
|
rand_data_t = Tensor(name='rand_data',
|
|
shape=(bs, ),
|
|
dtype=tensorrt_llm.torch_dtype_to_trt(
|
|
torch.float32))
|
|
|
|
outputs = tensorrt_llm.functional.categorical_sample(x, rand_data_t)
|
|
outputs.trt_tensor.name = 'output'
|
|
network.mark_output(outputs.trt_tensor)
|
|
# save onnx
|
|
# model_path = 'sample.onnx'
|
|
# to_onnx(net.trt_network, model_path)
|
|
|
|
# trt run
|
|
build_engine = EngineFromNetwork((builder.trt_builder, net.trt_network))
|
|
samples = []
|
|
nreps = 20000
|
|
with TrtRunner(build_engine) as runner:
|
|
for _ in range(nreps):
|
|
# NOTE: we need rand() here since TRT rand() produces same numbers
|
|
rand_data = torch.rand((bs, ), dtype=torch.float32)
|
|
outputs = runner.infer(feed_dict={
|
|
'x': probs.numpy(),
|
|
'rand_data': rand_data.numpy(),
|
|
})
|
|
# print(outputs)
|
|
samples.append(torch.tensor(outputs['output']))
|
|
# assert False, "PARTIAL"
|
|
samples = torch.stack(samples).float()
|
|
print(samples)
|
|
hist = []
|
|
bins = torch.arange(nbins + 1).float()
|
|
for i in range(bs):
|
|
h = torch.histogram(samples[:, i], bins=bins).hist
|
|
h = h / h.sum(-1)
|
|
hist.append(h)
|
|
print(hist)
|
|
|
|
# compare diff
|
|
np.testing.assert_allclose(torch.stack(hist), scaled_probs, atol=1e-2)
|
|
# assert False, "FORCED"
|
|
return
|