tvalentyn commented on code in PR #22131: URL: https://github.com/apache/beam/pull/22131#discussion_r940720337
########## sdks/python/apache_beam/ml/inference/tensorrt_inference.py: ########## @@ -0,0 +1,281 @@ +# +# Licensed to the Apache Software Foundation (ASF) under one or more +# contributor license agreements. See the NOTICE file distributed with +# this work for additional information regarding copyright ownership. +# The ASF licenses this file to You 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. +# + +# pytype: skip-file + +import logging +import sys +from typing import Any +from typing import Dict +from typing import Iterable +from typing import Optional +from typing import Sequence +from typing import Tuple + +import numpy as np + +import tensorrt as trt +from apache_beam.io.filesystems import FileSystems +from apache_beam.ml.inference.base import ModelHandler +from apache_beam.ml.inference.base import PredictionResult +from cuda import cuda + +TRT_LOGGER = trt.Logger(trt.Logger.INFO) + +logging.basicConfig(level=logging.INFO) +logging.getLogger("TensorRTEngineHandlerNumPy").setLevel(logging.INFO) +log = logging.getLogger("TensorRTEngineHandlerNumPy") + + +def _load_engine(engine_path): + file = FileSystems.open(engine_path, 'rb') + runtime = trt.Runtime(TRT_LOGGER) + engine = runtime.deserialize_cuda_engine(file.read()) + assert engine + return engine + + +def _load_onnx(onnx_path): + builder = trt.Builder(TRT_LOGGER) + network = builder.create_network( + flags=1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)) + parser = trt.OnnxParser(network, TRT_LOGGER) + with FileSystems.open(onnx_path) as f: + if not parser.parse(f.read()): + log.error("Failed to load ONNX file: %s", onnx_path) + for error in range(parser.num_errors): + log.error(parser.get_error(error)) + sys.exit(1) + return network, builder + + +def _build_engine(network, builder): + config = builder.create_builder_config() + runtime = trt.Runtime(TRT_LOGGER) + plan = builder.build_serialized_network(network, config) + engine = runtime.deserialize_cuda_engine(plan) + builder.reset() + return engine + + +def _validate_inference_args(inference_args): + """Confirms that inference_args is None. + + TensorRT engines do not need extra arguments in their execute_v2() call. + However, since inference_args is an argument in the RunInference interface, + we want to make sure it is not passed here in TensorRT's implementation of + RunInference. + """ + if inference_args: + raise ValueError( + 'inference_args were provided, but should be None because TensorRT ' + 'engines do not need extra arguments in their execute_v2() call.') + + +def ASSERT_DRV(args): + """CUDA error checking.""" + err, ret = args[0], args[1:] + if isinstance(err, cuda.CUresult): + if err != cuda.CUresult.CUDA_SUCCESS: + raise RuntimeError("Cuda Error: {}".format(err)) + else: + raise RuntimeError("Unknown error type: {}".format(err)) + # Special case so that no unpacking is needed at call-site. + if len(ret) == 1: + return ret[0] + return ret + + +class TensorRTEngine: + def __init__(self, engine: trt.ICudaEngine): + """Implementation of the TensorRTEngine class which handles + allocations associated with TensorRT engine. + + Example Usage:: + + TensorRTEngine(engine) + + Args: + engine: trt.ICudaEngine object that contains TensorRT engine + """ + self.engine = engine + self.context = engine.create_execution_context() + self.inputs = [] + self.outputs = [] + self.gpu_allocations = [] + self.cpu_allocations = [] + """Setup I/O bindings.""" + for i in range(self.engine.num_bindings): + name = self.engine.get_binding_name(i) + dtype = self.engine.get_binding_dtype(i) + shape = self.engine.get_binding_shape(i) + size = trt.volume(shape) * dtype.itemsize + allocation = ASSERT_DRV(cuda.cuMemAlloc(size)) + binding = { + 'index': i, + 'name': name, + 'dtype': np.dtype(trt.nptype(dtype)), + 'shape': list(shape), + 'allocation': allocation, + 'size': size + } + self.gpu_allocations.append(allocation) + if self.engine.binding_is_input(i): + self.inputs.append(binding) + else: + self.outputs.append(binding) + + assert self.context + assert len(self.inputs) > 0 + assert len(self.outputs) > 0 + assert len(self.gpu_allocations) > 0 + + for output in self.outputs: + self.cpu_allocations.append(np.zeros(output['shape'], output['dtype'])) + # Create CUDA Stream. + self.stream = ASSERT_DRV(cuda.cuStreamCreate(0)) + + def get_engine_attrs(self): + """Returns TensorRT engine attributes.""" + return ( + self.engine, + self.context, + self.inputs, + self.outputs, + self.gpu_allocations, + self.cpu_allocations, + self.stream) + + +class TensorRTEngineHandlerNumPy(ModelHandler[np.ndarray, + PredictionResult, + TensorRTEngine]): + def __init__(self, min_batch_size: int, max_batch_size: int, **kwargs): + """Implementation of the ModelHandler interface for TensorRT. + + Example Usage:: + + pcoll | RunInference( + TensorRTEngineHandlerNumPy( + min_batch_size=1, + max_batch_size=1, + engine_path="my_uri")) + + Args: + min_batch_size: minimum accepted batch size. + max_batch_size: maximum accepted batch size. + kwargs: Additional arguments like 'engine_path' and 'onnx_path' are + currently supported. + + See https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/ + for details + """ + self.min_batch_size = min_batch_size + self.max_batch_size = max_batch_size + if 'engine_path' in kwargs: + self.engine_path = kwargs.get('engine_path') + elif 'onnx_path' in kwargs: + self.onnx_path = kwargs.get('onnx_path') + + trt.init_libnvinfer_plugins(TRT_LOGGER, namespace="") + + def batch_elements_kwargs(self): + """Sets min_batch_size and max_batch_size of a TensorRT engine.""" + return { + 'min_batch_size': self.min_batch_size, + 'max_batch_size': self.max_batch_size + } + + def load_model(self) -> TensorRTEngine: + """Loads and initializes a TensorRT engine for processing.""" + engine = _load_engine(self.engine_path) + return TensorRTEngine(engine) + + def load_onnx(self) -> Tuple[trt.INetworkDefinition, trt.Builder]: + """Loads and parses an onnx model for processing.""" + return _load_onnx(self.onnx_path) + + def build_engine( + self, network: trt.INetworkDefinition, + builder: trt.Builder) -> TensorRTEngine: + """Build an engine according to parsed/created network.""" + engine = _build_engine(network, builder) + return TensorRTEngine(engine) + + def run_inference( Review Comment: > We'd want to first resolve the thread-safety concern with 1 GPU. We need to address the thread safety concern per feedback in this thread. At minimum, let's add a lock and avoid sharing the context between the threads. I discussed this a bit with lcwik. In short term, we should implement a shared object cache for managing shared execution contexts. We should avoid the affinity between thread IDs and contexts, and instead make it possible to transfer shared contexts between threads. We should be able to rely on `start_bundle` and `finish_bundle` to claim and release the contexts from the shared pool. Open question is GPU memory management: - How many contexts do we create and how much GPU memory should we allocate per each context. Can we know at runtime what's the max # of context we can create for a particular engine? What are the tradeoffs between creating more contexts with less memory per contexts vs less contexts? Will pipeline degrade gracefully when more memory is needed or something will crash? If we solve memory management, adding the shared pool should not be too complex. We can refer to `UnboundedThreadPoolExecutor` implementation for inspiration. - In the future, we can consider sharing the object cache w/ contexts between multiple SDK processes (not just between threads) via `multiprocessing`. This may be worthwhile with A100 accelerators that have more GPU memory that can be shared across multiple processes, and this could allow users benefiting from MPS. We'd have to share contexts between processes - are these objects large in size? If it is just metadata, the overhead may not be significant. This is something we should also consider for Tensorflow runinference, since Tensorflow does not work well when using the same GPU from multiple processes and manual GPU memory management is needed, so Beam RunInference could help make it easier. In the medium term, we may need to recommend limiting pipelines to use 1 process only. Summary of questions asked above: - How many contexts can we create and how do you recommend to manage memory b/w them? - Are context objects small? - Does TensorRT support concurrent execution on multiple processes on the same GPU? I believe, I asked this and answer was `yes`, just, double checking... If yes, how GPU memory management is solved? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
