File size: 11,706 Bytes
5000658
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
# 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.
from __future__ import annotations

import contextlib
from dataclasses import dataclass
from typing import Any, Dict, List, Optional

# isort: off
import torch
import tensorrt as trt
# isort: on

from .._utils import torch_dtype_to_trt, trt_dtype_to_torch, trt_gte_10
from ..logger import logger


@contextlib.contextmanager
def _scoped_stream():
    '''Create a scoped cuda stream, and synchronize it when the context is destroyed
    '''
    #TODO: delete torch, use cuda native python bindings
    import torch
    stream = torch.cuda.current_stream()
    try:
        # return a handle, trt and other lib does not recognize torch.cuda.Stream
        yield stream.cuda_stream
    finally:
        stream.synchronize()


@dataclass
class TensorInfo:
    name: str
    dtype: trt.DataType
    shape: tuple
    # add more info like strides, formats if needed


class Session(object):
    ''' Session is a managed TensorRT runtime.  '''

    def __init__(self, **kwargs):
        # use Session.from_serialized_engine to create a session
        pass

    def _init(self, engine_buffer=None):
        '''
        @brief: Setup TensorRT engines and context from a serialized engine file
        @param engine_buffer: a buffer holds the serialized TRT engine
        '''
        self._runtime = trt.Runtime(logger.trt_logger)
        if engine_buffer is not None:
            self._engine = self.runtime.deserialize_cuda_engine(engine_buffer)

        self._context = None
        if not (trt_gte_10() and self.engine.streamable_weights_size):
            self.__prepare_execution_contexts()
        return self

    def __prepare_execution_contexts(self):
        self._context = self.engine.create_execution_context()
        assert self._context is not None, "Failed to create an execution context!"
        with _scoped_stream() as stream:
            self._context.set_optimization_profile_async(0, stream)

    @staticmethod
    def from_serialized_engine(engine) -> Session:
        '''
        @brief: Create a session from a serialized engine
        @param engine: a serialized engine
        @return: a Session object
        '''
        session = Session()
        return session._init(engine)

    @staticmethod
    def from_engine(engine) -> Session:
        '''
        @brief: Create a session from an existing ICudaEngine engine
        @param engine: an ICudaEngine
        @return: a Session object
        '''
        session = Session()
        session.engine = engine
        return session._init()

    @property
    def runtime(self) -> trt.Runtime:
        return self._runtime

    @property
    def engine(self) -> trt.ICudaEngine:
        return self._engine

    @engine.setter
    def engine(self, engine: trt.ICudaEngine):
        self._engine = engine

    @property
    def context(self) -> trt.IExecutionContext:
        '''
        @brief: Get the default TensorRT execution context,
            use self.engine.create_execution_context() to create a new context if needed
        @return: one TensorRT execution context object
        '''
        return self._context

    @property
    def context_mem_size(self) -> int:
        return self.engine.device_memory_size

    def _print_engine_info(self):
        '''print engine info for debug purpose, internal use only.
        '''
        refittable = self.engine.refittable
        num_layers = self.engine.num_layers
        device_memory_size = self.engine.device_memory_size
        name = self.engine.name
        nb_profiles = self.engine.num_optimization_profiles
        logger.info(
            f"Engine:{name=:}, {refittable=:}, {num_layers=:}, {device_memory_size=:}, {nb_profiles=:}"
        )
        self._print_io_info()

    def _print_io_info(self):
        '''print engine i/o info for debug purpose, internal use only.
        '''

        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            mode = self.engine.get_tensor_mode(name)
            shape = self.engine.get_tensor_shape(name)
            dtype = self.engine.get_tensor_dtype(name)
            tformat = ";".join([
                self.engine.get_tensor_format_desc(name, p)
                for p in range(self.engine.num_optimization_profiles)
            ])
            logger.info(
                f"Tensor:{name=:}, {mode=:}, {shape=:}, {dtype=:}, {tformat=:}")

    def set_shapes(self,
                   tensor_dict: Dict[str, torch.Tensor],
                   context: Optional[trt.IExecutionContext] = None):
        if context is None:
            context = self.context

        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
                ok = context.set_input_shape(name, tensor_dict[name].shape)
                logger.debug(
                    f"setting input tensor {name} with shape {tensor_dict[name].shape}"
                )
                if not ok:
                    raise ValueError(
                        f"Couldn't assign {name} with shape {tensor_dict[name].shape}, "
                        f"engine supports [min, opt, max] = {self.engine.get_tensor_profile_shape(name, context.active_optimization_profile)}"
                    )

    def infer_shapes(
            self,
            inputs: List[TensorInfo],
            context: Optional[trt.IExecutionContext] = None
    ) -> List[TensorInfo]:
        '''
        @brief: Set input shapes to given context, and infer the output shapes from the given input shapes.
               This function should be called every time when the input shapes are changed before calling run().
               Or call the context.set_input_shape on all dynamic shaped input tensors manually.
        @param inputs: list of TensorInfo object, each item represents an input tensor
        @param context: TensorRT execution context, if None, use the default context
        @return: list of TensorInfo object, each item represents an output tensor, returns None if failed
        '''
        # set shape to the default context if context is not specified
        if context is None:
            context = self.context
        for i in inputs:
            if self.engine.get_tensor_mode(i.name) != trt.TensorIOMode.INPUT:
                raise ValueError(f"Tensor:{i.name} is not an input tensor")
            if self.engine.get_tensor_dtype(i.name) != i.dtype:
                raise ValueError(f"Tensor:{i.name} has wrong dtype")
            if not context.set_input_shape(i.name, i.shape):
                raise RuntimeError(
                    f"Could not set shape {i.shape} for tensor {i.name}. Please check the profile range for which your model was build."
                )

        outputs = []
        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
                shape = context.get_tensor_shape(name)
                dtype = self.engine.get_tensor_dtype(name)
                outputs.append(TensorInfo(name, dtype, shape))
        return outputs

    def _set_weight_streaming(self, gpu_weights_percent):
        assert self.engine is not None

        self._context = None

        if not trt_gte_10():
            assert gpu_weights_percent == 1, "Weight streaming is only supported by TensorRT 10.0 or later."
            return
        else:
            min = self.engine.minimum_weight_streaming_budget
            max = self.engine.streamable_weights_size
            budget = int(min + gpu_weights_percent * (max - min))

            budget_config = budget if gpu_weights_percent != 1 else 0
            self.engine.weight_streaming_budget = budget_config
            assert self.engine.weight_streaming_budget == budget_config, "Failed to set weight streaming budget!"
            logger.info(
                f"Set gpu weights percent to {gpu_weights_percent}, which is {budget} bytes. Valid range: {min} bytes ~ {max} bytes."
            )

        if self.engine.streamable_weights_size:
            try:
                self.__prepare_execution_contexts()
            except:
                free_mem = torch.cuda.mem_get_info()[0]
                if free_mem < budget:
                    raise torch.cuda.OutOfMemoryError(
                        f"Out of Memory: Memory budget is {budget} bytes but only {free_mem} bytes are available on the GPU."
                    )
                raise

    def run(self,
            inputs: Dict[str, Any],
            outputs: Dict[str, Any],
            stream,
            context=None) -> bool:
        '''
        @brief: Run the TensorRT engine with the given inputs and outputs
        @param inputs: dict of input tensors, key is tensor name, value is tensor pointer or torch tensor
        @param outputs: dict of output tensors, key is tensor name, value is tensor pointer or torch tensor
        @param stream: cuda stream to enqueue the TensorRT engine on
        @param context: TensorRT execution context, if None, use the default context
        @return: True if enqueue succeeded, note the enqueue is an async call,
            returning True does not mean the execution is finished
        '''
        # enqueue to the default context if context is not specified
        if context is None:
            context = self.context

        import torch
        for tensor_name in inputs:
            tensor = inputs[tensor_name]
            ptr = tensor.data_ptr() if isinstance(tensor,
                                                  torch.Tensor) else tensor
            context.set_tensor_address(tensor_name, ptr)
        for tensor_name in outputs:
            tensor = outputs[tensor_name]
            ptr = tensor.data_ptr() if isinstance(tensor,
                                                  torch.Tensor) else tensor
            context.set_tensor_address(tensor_name, ptr)
        ok = context.execute_async_v3(stream)
        return ok

    def _debug_run(self,
                   inputs: Dict[str, "torch.Tensor"],
                   context=None) -> Dict[str, "torch.Tensor"]:
        '''Run the engine enqueue with allocated output tensors, for debug purpose, since it is a sync call and slower than run
        '''
        import torch

        inputs_info = [
            TensorInfo(name, torch_dtype_to_trt(tensor.dtype), tensor.shape)
            for name, tensor in inputs.items()
        ]
        outputs_info = self.infer_shapes(inputs_info)
        outputs = {
            t.name: torch.empty(tuple(t.shape),
                                dtype=trt_dtype_to_torch(t.dtype),
                                device='cuda')
            for t in outputs_info
        }
        with _scoped_stream() as stream:
            self.run(inputs=inputs,
                     outputs=outputs,
                     stream=stream,
                     context=context)
        return outputs