# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. # # 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 multiprocessing import os import sys import warnings from . import framework from .framework import _get_paddle_place, _get_paddle_place_list from .framework import cuda_places, cpu_places, xpu_places from . import core __all__ = [ 'CompiledProgram', 'ExecutionStrategy', 'BuildStrategy', 'IpuCompiledProgram', 'IpuStrategy', ] ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy BuildStrategy = core.ParallelExecutor.BuildStrategy InferNativeConfig = core.NativeConfig InferAnalysisConfig = core.AnalysisConfig DeviceType = core.DeviceType def _place_obj(place): p = core.Place() p.set_place(place) return p def _is_pserver_mode(main_program): main = main_program if main_program else framework.default_main_program() for op in main.global_block().ops: if op.type in ["send", "recv"]: return True return False def _has_backward_op(graph): for node in graph.nodes(): if ( node.is_op() and node.op() is not None and node.op().type().endswith("_grad") ): return True return False def _prune_feed_ops(program): # prune the feed ops in the program. pop_idx = [] for i, op in enumerate(program.global_block().ops): if op.type == "feed": pop_idx.append(i) for index in pop_idx[::-1]: program.global_block()._remove_op(index) def _has_optimize_op(block): for op in block.ops: op_maker = core.op_proto_and_checker_maker optimize = core.op_proto_and_checker_maker.OpRole.Optimize if op_maker.kOpRoleVarAttrName() in op.attr_names and int( op.all_attrs()[op_maker.kOpRoleAttrName()] ) == int(optimize): return True return False def _has_optimizer_in_control_flow(program): if not program: program = framework.default_main_program() for op in program.global_block().ops: if op.type == "conditional_block_grad": sub_block = program.block(op._block_attr_id("sub_block")) if _has_optimize_op(sub_block): return True return False def _should_broadcast_or_not_exists(program, var_name): block = program.global_block() var = block.vars.get(var_name, None) if var is None: return True is_distributed = getattr(var, '_is_distributed', False) or getattr( var, 'is_distributed', False ) return not is_distributed class CompiledProgram: """ :api_attr: Static Graph The CompiledProgram is used to transform a program or graph for various optimizations according to the configuration of build_strategy, for example, the operators' fusion in the computation graph, memory optimization during the execution of the computation graph, etc. For more information about build_strategy, please refer to :code:`paddle.static.BuildStrategy`. Args: program_or_graph (Graph|Program): This argument is the Program or Graph being executed. build_strategy(BuildStrategy): This argument is used to compile the program or graph with the specified options, such as operators' fusion in the computational graph and memory optimization during the execution of the computational graph. For more information about build_strategy, please refer to :code:`paddle.static.BuildStrategy`. The default is None. Returns: CompiledProgram Example: .. code-block:: python import numpy import paddle import paddle.static as static paddle.enable_static() place = paddle.CUDAPlace(0) # paddle.CPUPlace() exe = static.Executor(place) data = static.data(name='X', shape=[None, 1], dtype='float32') hidden = static.nn.fc(x=data, size=10) loss = paddle.mean(hidden) paddle.optimizer.SGD(learning_rate=0.01).minimize(loss) exe.run(static.default_startup_program()) compiled_prog = static.CompiledProgram( static.default_main_program()) x = numpy.random.random(size=(10, 1)).astype('float32') loss_data, = exe.run(compiled_prog, feed={"X": x}, fetch_list=[loss.name]) """ def __init__(self, program_or_graph, build_strategy=None): if isinstance(program_or_graph, core.Graph): self._graph = program_or_graph # don't not create a new program here. self._program = None elif isinstance(program_or_graph, framework.Program): _prune_feed_ops(program_or_graph) self._graph = core.Graph(program_or_graph.desc) self._program = program_or_graph else: raise TypeError( "The type of program_to_graph parameter is wrong, expected Graph or Program, but received %s" % type(program_or_graph) ) self._scope = None self._place = None self._executor = None self._compiled = False self._is_data_parallel = False self._is_inference = False self._loss_name = None self._share_vars_from = None self._places = None self._build_strategy = build_strategy self._exec_strategy = None def with_data_parallel( self, loss_name=None, build_strategy=None, exec_strategy=None, share_vars_from=None, places=None, ): """ This interface is used to transform the input Program or Graph to a multi-graph to run the model in data parallel mode. Users can use the build_strategy and exec_strategy to set some optimizations that can be applied during the construction and computation of the Graph, such as reducing the number of AllReduce operations, specifying the size of the thread pool used in the computation Graph running the model, and so on. .. note:: If build_strategy is specified when building CompiledProgram and calling with_data_parallel, build_strategy in CompiledProgram will be overwritten, therefore, if it is data parallel training, it is recommended to set build_strategy when calling with_data_parallel interface. Args: loss_name (str): This parameter is the name of the loss Tensor of the model. **Note: If it is model training, you must set loss_name, otherwise the result may be problematic**. The default is None. build_strategy(BuildStrategy): This parameter is used to compile the program or graph with the specified options, such as operators' fusion in the computational graph and memory optimization during the execution of the computational graph. For more information about build_strategy, please refer to :code:`fluid.BuildStrategy`. The default is None. exec_strategy(ExecutionStrategy): exec_strategy specifies the options that can be changed when running the current model, such as the thread pool size. For more information about exec_strategy, please refer to :code:`fluid.ExecutionStrategy`. The default is None. share_vars_from(CompiledProgram): If share_vars_from is set, the current CompiledProgram will share the parameter value with the CompiledProgram specified by share_vars_from. This parameter needs to be set when model testing is required during model training, and the data parallel mode is used for training and testing. Since CompiledProgram will only distribute parameter Tensors to other devices when it is first executed, the CompiledProgram specified by share_vars_from must be run before the current CompiledProgram. The default is None. places(list(CUDAPlace)|list(CPUPlace)|list(str)|None): This parameter specifies the device on which the model is running. If you want to run on GPU0 and GPU1, places are [fluid.CUDAPlace(0), fluid.CUDAPlace(1)]; if you want to run with 2 CPUs, places are [fluid.CPUPlace()] * 2. If the parameter is not set, i.e. the parameter is None, the available device will be obtained from the environment variable when the model is executed: If the GPU is used, the currently available device ID is obtained from the environment variable FLAGS_selected_gpus or CUDA_VISIBLE_DEVICES when the model is executed; CPU, when the model is executed, the currently available CPU number is obtained from the environment variable CPU_NUM. For example, export CPU_NUM=4, if the environment variable is not set, the executor will add the variable to the environment variable and set its value to 1. The default is None. If ``places`` is the list of string, the string in the list can be ``cpu``, ``gpu:x``, where ``x`` is the index of the GPUs. Returns: CompiledProgram Example: .. code-block:: python import numpy import os import paddle import paddle.static as static paddle.enable_static() use_cuda = True place = paddle.CUDAPlace(0) if use_cuda else paddle.CPUPlace() parallel_places = [paddle.CUDAPlace(0), paddle.CUDAPlace(1)] if use_cuda else [paddle.CPUPlace()] * 2 # NOTE: If you use CPU to run the program, you need # to specify the CPU_NUM, otherwise, paddle will use # all the number of the logic core as the CPU_NUM, # in that case, the batch size of the input should be # greater than CPU_NUM, if not, the process will be # failed by an exception. if not use_cuda: os.environ['CPU_NUM'] = str(2) exe = static.Executor(place) data = static.data(name='X', shape=[None, 1], dtype='float32') hidden = static.nn.fc(x=data, size=10) loss = paddle.mean(hidden) test_program = static.default_main_program().clone(for_test=True) paddle.optimizer.SGD(learning_rate=0.01).minimize(loss) exe.run(static.default_startup_program()) compiled_train_prog = static.CompiledProgram( static.default_main_program()).with_data_parallel( loss_name=loss.name, places=parallel_places) # NOTE: if not set share_vars_from=compiled_train_prog, # the parameters used in test process are different with # the parameters used by train process compiled_test_prog = static.CompiledProgram( test_program).with_data_parallel( share_vars_from=compiled_train_prog, places=parallel_places) train_data = numpy.random.random(size=(10, 1)).astype('float32') loss_data, = exe.run(compiled_train_prog, feed={"X": train_data}, fetch_list=[loss.name]) test_data = numpy.random.random(size=(10, 1)).astype('float32') loss_data, = exe.run(compiled_test_prog, feed={"X": test_data}, fetch_list=[loss.name]) """ assert ( not self._is_data_parallel ), "Already compiled with parallel, cannot be recompiled." assert ( not self._is_inference ), "Cannot compile with both data parallel and inference." self._is_data_parallel = True # FIXME(zcd): Currently, the build_strategy can be set during creating # CompiledProgram or calling with_data_parallel, and it may be confusing, # but in the long run, we should set up build_strategy only when creating # CompiledProgram, and exec_strategy should be deprecated. if build_strategy is not None: self._build_strategy = build_strategy self._exec_strategy = exec_strategy self._loss_name = loss_name self._share_vars_from = share_vars_from if isinstance(places, (list, tuple)): self._places = _get_paddle_place_list(places) else: self._places = _get_paddle_place(places) if _has_backward_op(self._graph): assert ( self._loss_name is not None ), "The loss name of CompiledProgram is None. The loss name should be set if CompiledProgram contains backward part." if self._places is not None: if not isinstance(self._places, (list, tuple)): self._places = [self._places] return self def _with_inference_optimize(self, config): """Add inference optimize Args: config: instance of `NativeConfig` or `AnalysisConfig` to create predictor Returns: self """ assert ( not self._is_data_parallel ), "Cannot compile with both data parallel and inference" assert ( not self._is_inference ), "Already compiled with inference, cannot be recompiled." assert any( [ isinstance(config, InferNativeConfig), isinstance(config, InferAnalysisConfig), ] ) self._is_inference = True self._infer_config = config return self def _with_distributed(self): raise NotImplementedError( "Subclass of CompiledProgram should implement _with_distributed method." ) def _compile_data_parallel(self, places, use_device, scope=None): if self._share_vars_from: if scope: sys.stderr.write("share_vars_from is set, scope is ignored.\n") if not self._share_vars_from._is_data_parallel: raise ValueError( "The shared Program is not data parallel, cannot " "share variables from it." ) if self._share_vars_from._executor is None: raise ValueError( "The shared Program is not compiled and executed, so there is no " "variables to share." ) self._local_scopes = self._share_vars_from._executor.local_scopes() else: assert scope is not None, "" self._local_scopes = [] assert isinstance(places, tuple) or isinstance( places, list ), "Currently , The places type can only be list or tuple, but the input type is {}.".format( type(places) ) if self._build_strategy is None: self._build_strategy = BuildStrategy() self._build_strategy.is_distribution = _is_pserver_mode(self._program) if self._exec_strategy is None: self._exec_strategy = ExecutionStrategy() self._exec_strategy._use_device = use_device if self._exec_strategy.num_threads == 0: if self._exec_strategy._use_device == DeviceType.CUDA: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. self._exec_strategy.num_threads = len(places) * 4 elif self._exec_strategy._use_device == DeviceType.XPU: # Currently only single thread is supported in Kunlun XPU. self._exec_strategy.num_threads = 1 else: self._exec_strategy.num_threads = len(places) * 2 if ( "FLAGS_use_cinn" in core.globals() and core.globals()["FLAGS_use_cinn"] and self._exec_strategy.num_threads != 1 ): warnings.warn( "At present, when CINN is turned on, each process can " "only contain one thread, so reset the number of threads to 1 here." ) self._exec_strategy.num_threads = 1 if self._build_strategy.num_trainers > 1: assert self._is_data_parallel, ( "If you use multi-trainer to train the model, you should use " "the data parallel model, i.e. calling with_data_parallel function." ) # TODO(wuyi): trainer endpoings should be passed in through # build_strategy, not program.xxx. # TODO(gongwb): let user to set them once. if ( self._program and self._build_strategy.num_trainers > 1 and self._program._trainers_endpoints ): tps = self._program._trainers_endpoints assert self._build_strategy.num_trainers == len( tps ), "The trainer numbers is not equal to endpoint numbers." self._build_strategy.trainers_endpoints = tps if self._program: self._build_strategy.nccl_comm_num = self._program._nccl_comm_num self._build_strategy.use_hierarchical_allreduce = ( self._program._use_hierarchical_allreduce ) self._build_strategy.hierarchical_allreduce_inter_nranks = ( self._program._hierarchical_allreduce_inter_nranks ) if self._build_strategy.sync_batch_norm: self._build_strategy.enable_sequential_execution = True if self._program is not None and self._program._enable_dgc: assert ( self._exec_strategy._use_device == DeviceType.CUDA ), "DGC only used under CUDA environment." assert ( self._build_strategy.num_trainers * len(places) > 1 ), "DGC is not avaliable for single card training." assert ( self._build_strategy.reduce_strategy == BuildStrategy.ReduceStrategy.AllReduce ), "DGC \ only can be used for AllReduce BuildStrategy." # DGC doesn't support fuse for now, close fuse. self._build_strategy.fuse_all_reduce_ops = False self._persistable_vars = [] for node in self._graph.nodes(): if ( node.is_var() and node.var() is not None and node.var().persistable() and node.var().type() != core.VarDesc.VarType.RAW ): name = node.name() if ( self._program is not None and _should_broadcast_or_not_exists(self._program, name) ): self._persistable_vars.append(node.name()) places = list(map(_place_obj, places)) # ParallelExecutor would broadcast all the parameters during initializing. # The parameters of each process should be in the same ordered for the data-parallelism # distributed training to keep the broadcast correct. self._persistable_vars = list(set(self._persistable_vars)) self._persistable_vars.sort() return core.ParallelExecutor( places, self._persistable_vars, self._loss_name if self._loss_name else '', self._scope, self._local_scopes, self._exec_strategy, self._build_strategy, self._graph, ) def _compile_inference(self): return core.create_paddle_predictor(self._infer_config) def _compile(self, scope, place): """Compile the program based on the configs. Args: scope: The variables (resources) that are associated with this compiled program. place: The location that the compiled program will be run on. Returns: self """ if self._compiled: if scope and self._scope != scope: raise ValueError("Cannot compile program with different scope.") if place and not self._place._equals(place): raise ValueError("Cannot compile program with different place.") return self self._compiled = True self._scope = scope self._place = place if self._is_inference: self._executor = self._compile_inference() else: if self._is_data_parallel: self._places = self._get_places(self._place, self._places) else: self._places = [self._place] # Todo(liym27):If optimizer is used in control flow, # training on multi-places is not supported now, will # be supported later. if len(self._places) > 1 and _has_optimizer_in_control_flow( self._program ): raise NotImplementedError( "If optimizer is used in control flow, " "training on multi-places is not supported now." ) if isinstance(self._place, core.CUDAPlace): use_device = DeviceType.CUDA elif isinstance(self._place, core.XPUPlace): use_device = DeviceType.XPU else: use_device = DeviceType.CPU self._executor = self._compile_data_parallel( use_device=use_device, scope=self._scope, places=self._places ) return self def _get_places(self, place, place_list): has_set_place = place_list is not None if has_set_place: for p in place_list: assert ( p._type() == place._type() ), "Place type not match. You may set wrong type of places." else: if isinstance(place, core.CUDAPlace): place_list = cuda_places() elif isinstance(place, core.XPUPlace): place_list = xpu_places() else: place_list = cpu_places() assert place_list, "No places for execution." return place_list class IpuDynamicPatcher: """ Patcher for IPU dynamic2static support. """ patcher_cache = [] def __init__(self): pass @staticmethod def convert_concrete_program( ipu_strategy, concrete_program, class_instance=None ): """ Convert the ConcreteProgram to IPUConcreteProgram. """ from ..fluid.dygraph.base import switch_to_static_graph from ..fluid import backward from ..fluid.initializer import Constant from ..fluid.framework import device_guard import paddle inputs = concrete_program.inputs outputs = concrete_program.outputs startup_program = concrete_program.startup_program scope = paddle.static.global_scope() @switch_to_static_graph def append_backward_desc(): program = concrete_program.main_program # backward with optimizer to add backward graph to program backward.gradients_with_optimizer(program, ipu_strategy._optimizer) # initialize backward parameters exe = paddle.static.Executor(paddle.CPUPlace()) startup_program = paddle.static.default_startup_program() exe.run(startup_program) return program if ipu_strategy.enable_fp16: class_instance.to(dtype="float16") # copy the bias and filters for param_or_buffer in concrete_program.parameters: param_or_buffer_tensor = scope.var( param_or_buffer.name ).get_tensor() src_tensor = param_or_buffer.value().get_tensor() param_or_buffer_tensor._share_data_with(src_tensor) # TODO(czr): feed and fetch list needs to consider more type if class_instance: feed_list = [elem.name for elem in inputs[1:] if elem is not None] else: feed_list = [elem.name for elem in inputs if elem is not None] fetch_list = [elem.name for elem in outputs] if ipu_strategy.is_training: concrete_program.main_program = append_backward_desc() # copy optimizer parameters optimizer = ipu_strategy._optimizer for k, v in optimizer._accumulators.items(): for param_name, var_tmp in v.items(): var = optimizer.helper.create_global_variable( name=var_tmp.name, persistable=True, dtype=var_tmp.dtype, type=var_tmp.type, shape=var_tmp.shape, belong_to_optimizer=True, ) device = optimizer._get_device_for_param(param_name) with device_guard(device): optimizer.helper.set_variable_initializer( var, initializer=Constant(value=0.0) ) param_or_lr_tensor = scope.find_var( var_tmp.name ).get_tensor() optim_tensor = var.value().get_tensor() param_or_lr_tensor._share_data_with(optim_tensor) optimizer._accumulators[k][param_name] = var @switch_to_static_graph def func_compile(): if ipu_strategy.enable_fp16: amp_list = paddle.static.amp.CustomOpLists() amp_list.unsupported_list = {"cumsum"} to_fp16_var_names = paddle.static.amp.cast_model_to_fp16( concrete_program.main_program, amp_list, use_fp16_guard=False, ) paddle.static.amp.cast_parameters_to_fp16( paddle.CPUPlace(), concrete_program.main_program, to_fp16_var_names=to_fp16_var_names, ) program = IpuCompiledProgram( concrete_program.main_program, ipu_strategy=ipu_strategy, scope=scope, ).compile(feed_list, fetch_list) return program main_program = func_compile() concrete_program.main_program = main_program return concrete_program @staticmethod def patch_program_cache(ipu_strategy): """Monkey patch ProgramCache discriptor to support dynamic2static in IPU. Args: ipu_strategy: The ipu_strategy used in dynamic graph. Returns: None """ from paddle.jit.dy2static.program_translator import ( CacheKey, ProgramCache, MAX_TRACED_PROGRAM_COUNT, ) from paddle.jit.dy2static import logging_utils from paddle.jit.dy2static.partial_program import ( partial_program_from, ) old_getter = ProgramCache.__getitem__ def patch_getter(self, item): if not isinstance(item, CacheKey): raise ValueError( 'type(item) should be CacheKey, but received %s' % type(item).__name__ ) item_id = hash(item) self._recent_key = item_id if item_id not in self._caches or ipu_strategy.need_compile: if item_id in self._caches: logging_utils.warn( "ipu_strategy chances detected. Please sync weights." ) if self._caches and not ipu_strategy.need_compile: logging_utils.warn( "dynamic2static on IPU doesn't support mutiple caches. Please make sure" "dynamic inputs is not used." ) concrete_program, _ = self._build_once(item) concrete_program = IpuDynamicPatcher.convert_concrete_program( ipu_strategy, concrete_program, item.class_instance ) self._caches[item_id] = ( concrete_program, partial_program_from(concrete_program), ) # Note: raise warnings if number of traced program is more than `max_tracing_count` current_tracing_count = len(self._caches) if current_tracing_count > MAX_TRACED_PROGRAM_COUNT: logging_utils.warn( "Current traced program number: {} > `max_tracing_count`:{}. Too much cached programs will bring expensive overhead. " "The reason may be: (1) passing tensors with different shapes, (2) passing python objects instead of tensors.".format( current_tracing_count, MAX_TRACED_PROGRAM_COUNT ) ) return self._caches[item_id] setattr(ProgramCache, '__getitem__', patch_getter) IpuDynamicPatcher.patcher_cache.append( [ProgramCache, '__getitem__', old_getter] ) @staticmethod def patch_lr_scheduler(ipu_strategy): from paddle.optimizer.lr import LRScheduler # For IPU dynamic graph usage, lr_var is not synced in executor as static mode do. # Manually set lr to ipu_strategy to update the lr. old_step = LRScheduler.step def patch_step(self, epoch=None): old_step(self, epoch) ipu_strategy.set_options({"lr": self.last_lr}) setattr(LRScheduler, 'step', patch_step) IpuDynamicPatcher.patcher_cache.append([LRScheduler, 'step', old_step]) @staticmethod def register_patch(ipu_strategy): IpuDynamicPatcher.patch_program_cache(ipu_strategy) IpuDynamicPatcher.patch_lr_scheduler(ipu_strategy) @staticmethod def release_patch(): for module, key, attr in IpuDynamicPatcher.patcher_cache: setattr(module, key, attr) class IpuStrategy: """ Help users precisely control the graph building in :code:`paddle.static.IpuCompiledProgram` . Returns: The IpuStrategy instance. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() """ def __init__(self): if core.is_compiled_with_ipu(): self._ipu_strategy = core.IpuStrategy() default_options = { 'location_optimizer': { 'on_chip': 0, 'use_replicated_tensor_sharding': 1, }, # set optimizer location 'accumulation_and_replication_reduction_type': 1, # popart::ReductionType::Mean 'mean_accumulation_and_replication_reduction_strategy': 1, # popart::MeanReductionStrategy::Post } self._ipu_strategy.set_options(default_options) self.has_custom_ops = False self.custom_op_names = [] self.need_compile = True else: raise RuntimeError( "Can not use IpuStrategy in non IPU compiled environment, please re-compile with WITH_IPU=ON." ) from paddle import in_dynamic_mode if in_dynamic_mode(): self.register_patch() def register_patch(self): """ Register patchs function to support dynamic to static on IPU. This operation would break the dy2static functionality on CPU. Use `release_patch` to release the patch. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static ipu_strategy = static.IpuStrategy() ipu_strategy.register_patch() """ IpuDynamicPatcher.register_patch(self) def release_patch(self): """ Release the registered IPU functions. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static ipu_strategy = static.IpuStrategy() ipu_strategy.release_patch() """ IpuDynamicPatcher.release_patch() def set_optimizer(self, optimizer): """ Set optimizer to ipu_strategy in dynamic mode. Args: optimizer (Optimizer): Optimizer to be used in training. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static linear = paddle.nn.Linear(10, 10) optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=linear.parameters()) ipu_strategy = static.IpuStrategy() ipu_strategy.set_optimizer(optimizer) """ from paddle import in_dynamic_mode if in_dynamic_mode(): self._optimizer = optimizer optimizer_attrs = self.parse_optimizer(optimizer) self._ipu_strategy.set_options(optimizer_attrs) else: raise RuntimeError("Only needs to set optimizer in dynamic mode.") def parse_optimizer(self, optimizer): """ Parse optimizer attributes for IPU dynamic to static support. Currently only support parse lr. Args: optimizer (Optimizer): Optimizer to be parsed. Returns: Dict. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static linear = paddle.nn.Linear(10, 10) optimizer = paddle.optimizer.SGD(learning_rate=0.01, parameters=linear.parameters()) ipu_strategy = static.IpuStrategy() attrs = ipu_strategy.parse_optimizer(optimizer) """ def get_lr(): from paddle.optimizer.lr import LRScheduler if isinstance(optimizer._learning_rate, float): return {"lr": optimizer._learning_rate} elif isinstance(optimizer._learning_rate, LRScheduler): return {"lr": optimizer._learning_rate()} attr_fn = [get_lr] optimizer_attrs = {"is_dynamic": True} for fn in attr_fn: optimizer_attrs.update(fn()) return optimizer_attrs def set_graph_config( self, num_ipus=1, is_training=True, micro_batch_size=1, enable_manual_shard=False, ): """ Set graph configuration to the IpuStrategy instance. Args: num_ipus (int, optional): Number of IPU devices. Default 1, which means only use 1 IPU. is_training (bool, optional): True is training graph, False is inference graph. Default True, which means is training mode. batch_size (int, optional): The batch-size in the graph. Used to make the graph batch-size fixed, if the batch-size in the graph is dynamic. Default 1, which means the batch-size would be set 1, if the batch-size is dynamice. enable_manual_shard (bool, optional): Enable graph sharding or not. Only if num_ipus > 1, enable_manual_shard is able to be set True. Default False, which means disabled. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.set_graph_config(num_ipus=1, is_training=True, micro_batch_size=1, enable_manual_shard=False) """ if num_ipus == 1 and enable_manual_shard: raise RuntimeError( "Only if num_ipus > 1, enable_manual_shard is able to be set True." ) options = { 'num_ipus': num_ipus, 'is_training': is_training, 'micro_batch_size': micro_batch_size, 'enable_manual_shard': enable_manual_shard, } self.set_options(options) def set_pipelining_config( self, enable_pipelining=False, batches_per_step=1, enable_gradient_accumulation=False, accumulation_factor=1, ): """ Set pipelining configuration to the IpuStrategy instance. Used to optimize the throughput performance. Args: enable_pipelining (bool, optional): Enable data pipelining between subgraphs. Only if enable_manual_shard=True, enable_pipelining is able to be set True. Default False, which means disabled. batches_per_step (int, optional): Set the batches per run in data pipelining mode. Only if enable_pipelining=True, batches_per_step is able to be set > 1. Default 1, which means no data pipelining. enable_gradient_accumulation (bool, optional): Enable to accumulate gradients before updating the weights in training mode. Only if enable_pipelining=True, enable_gradient_accumulation is able to be set True. Default False, which means no gradient accumulation. accumulation_factor (int, optional): Specify the number of micro-batches to accumulate before applying the varUpdate. Default 1, which means disable the accumulation. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.set_pipelining_config(enable_pipelining=False, batches_per_step=1, enable_gradient_accumulation=False, accumulation_factor=1) """ enable_manual_shard = self.get_option('enable_manual_shard') if not enable_manual_shard and enable_pipelining: raise RuntimeError( "Only if enable_manual_shard=True, enable_pipelining is able to be set True." ) options = { 'enable_pipelining': enable_pipelining, 'batches_per_step': batches_per_step, 'enable_gradient_accumulation': enable_gradient_accumulation, 'accumulation_factor': accumulation_factor, } self.set_options(options) def set_precision_config(self, enable_fp16=False): """ Set half computation configuration to the IpuStrategy instance. Used to optimize the performance. Args: enable_fp16 (bool, optional): Enable FLOAT16 mode and transform FLOAT32 to FLOAT16. Default False, which means disable FLOAT16 mode. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.set_precision_config(enable_fp16=False) """ options = { 'enable_fp16': enable_fp16, } self.set_options(options) def add_custom_op( self, paddle_op, popart_op=None, domain='custom.ops', version=1 ): """ Add a mapping to use popart custom ops running on the IPU. Args: paddle_op(str): the name of custom op in paddle. popart_op(str): the name of custom op in popart. domain(str): domain name of custom op in popart. version(int): version of custom op in popart. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.add_custom_op('paddle_relu', 'popart_relu') """ if popart_op is None: popart_op = paddle_op custom_op = { 'paddle_op': paddle_op, 'popart_op': popart_op, 'domain': domain, 'version': version, } self.set_options({'custom_op': custom_op}) self.custom_op_names.append(paddle_op) if not self.has_custom_ops: self.has_custom_ops = True def set_options(self, options): """ Set options from dict. Args: options(dict): dict of options. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() options = {'num_ipus':1, 'enable_fp16': True} ipu_strategy.set_options(options) """ self._ipu_strategy.set_options(options) # check whether to recompile program with updated ipu options. recompile_white_list = {'lr'} if options.keys() - recompile_white_list: self.need_compile = True def get_option(self, option): """ Get option. Args: option(str): name of option. Returns: option value. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() num_ipus = ipu_strategy.get_option('num_ipus') """ return self._ipu_strategy.get_option(option)['value'] def enable_pattern(self, pattern): """ Enable PopART pattern to optimize the graph. Args: pattern(string): the name of the pattern. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.enable_pattern("ViewSimplifyPattern") """ self._ipu_strategy.enable_pattern(pattern) def disable_pattern(self, pattern): """ Disable PopART pattern. Args: pattern(string): the name of the pattern. Returns: None. Examples: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() ipu_strategy = static.IpuStrategy() ipu_strategy.disable_pattern("ViewSimplifyPattern") """ self._ipu_strategy.disable_pattern(pattern) @property def num_ipus(self): """ Get the number of IPU devices from IpuStrategy instance. """ return self.get_option('num_ipus') @property def is_training(self): """ Get the boolean of training or inference from IpuStrategy instance. """ return self.get_option('is_training') @property def enable_pipelining(self): """ Get the boolean of enable pipelining or not from IpuStrategy instance. """ return self.get_option('enable_pipelining') @property def enable_fp16(self): """ Get the boolean of float16 mode or not from IpuStrategy instance. """ return self.get_option('enable_fp16') class IpuCompiledProgram: """ The IpuCompiledProgram is used to transform a program to a ipu-target program, such as forward graph extraction, computing graph transformation, useless scale Ops clean, etc. Args: program(Program, optional): This parameter represents the :code:`Program` to be executed. Default is None, which means the program will be set to the default program :code:`paddle.static.default_main_program()` . scope(Scope, optional): The scope used to run this program, you can switch it to different scope. Default is None, which means use the global scope :code:`paddle.static.global_scope()` . ipu_strategy(IpuStrategy, optional): This argument is used to build the program with the specified options, such as half computation, training or inference session, the number of IPUs, etc. Default is None, which means build the program based on the default `ipu_strategy`. Returns: IpuCompiledProgram Example: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() a = static.data(name='data', shape=[None, 1], dtype='int32') b = a + 1 main_prog = static.default_main_program() ipu_strategy = static.IpuStrategy() ipu_strategy.set_graph_config(num_ipus=1, is_training=True, micro_batch_size=1) ipu_strategy.set_pipelining_config(enable_pipelining=False, batches_per_step=1, enable_gradient_accumulation=False, accumulation_factor=1) ipu_strategy.set_precision_config(enable_fp16=False) ipu_compiled_program = static.IpuCompiledProgram( main_prog, ipu_strategy=ipu_strategy) """ def __init__(self, program=None, scope=None, ipu_strategy=None): if not core.is_compiled_with_ipu(): raise ValueError( "Can not use this function since PaddlePaddle is not compiled with IPU" ) if program is None: program = framework.default_main_program() if not isinstance(program, framework.Program): raise TypeError( "The type of program is wrong, expected Program, but got %s" % type(program) ) self._program = program self._compiled = False if scope is not None: self._scope = scope else: # import here to avoiding confused import paddle self._scope = paddle.static.global_scope() if ipu_strategy is not None: self._ipu_strategy = ipu_strategy else: self._ipu_strategy = IpuStrategy() if ipu_strategy.has_custom_ops: self._custom_op_names = set(ipu_strategy.custom_op_names) else: self._custom_op_names = () self._backend = core.IpuBackend.get_instance() def compile(self, feed_list, fetch_list): """ This interface is used to compile the input Program to a program to run the model on the ipu. Args: feed_list(list): This parameter represents the input Tensors of the model. fetch_list(list): This parameter represents the Tensors that need to be returned after the model. Returns: Program Example: .. code-block:: python # required: ipu import paddle import paddle.static as static paddle.enable_static() a = static.data(name='data', shape=[None, 1], dtype='int32') b = a + 1 main_prog = static.default_main_program() ipu_strategy = static.IpuStrategy() ipu_strategy.set_graph_config(num_ipus=1, is_training=True, micro_batch_size=1) ipu_strategy.set_pipelining_config(enable_pipelining=False, batches_per_step=1, enable_gradient_accumulation=False, accumulation_factor=1) ipu_strategy.set_precision_config(enable_fp16=False) program = static.IpuCompiledProgram( main_prog, ipu_strategy=ipu_strategy).compile([a.name], [b.name]) """ self._backend.set_scope(self._scope) self._backend.set_ipu_strategy(self._ipu_strategy._ipu_strategy) # feed and fetch doesn't have corresponding popart op, so we rm both here global_block = self._program.global_block() need_to_remove_op_index = [] for i, op in enumerate(global_block.ops): op.desc.set_is_target(False) if op.type == 'feed' or op.type == 'fetch': need_to_remove_op_index.append(i) for index in need_to_remove_op_index[::-1]: global_block._remove_op(index) for var in ['feed', 'fetch']: if global_block.has_var(var): global_block._remove_var(var) self._program.desc.flush() self._graph = core.Graph(self._program.desc) if self._ipu_strategy.is_training: passes = [ 'optimizer_extract_pass', 'optimizer_state_align_pass', ] for pass_name in passes: a_pass = core.get_pass(pass_name) a_pass.apply(self._graph) passes = [ 'forward_graph_extract_pass', 'infer_shape_pass', 'avg_shard_pass', 'delete_scale_op_pass', ] for pass_name in passes: a_pass = core.get_pass(pass_name) if pass_name == 'infer_shape_pass': a_pass.set('feed_list', feed_list) a_pass.apply(self._graph) a_pass = core.get_pass('popart_canonicalization_pass') if self._custom_op_names: a_pass.set('custom_ops', self._custom_op_names) a_pass.apply(self._graph) passes = [ 'ipu_inplace_pass', 'ipu_graph_builder_pass', 'ipu_runtime_replacer_pass', ] for pass_name in passes: a_pass = core.get_pass(pass_name) a_pass.set('feed_list', feed_list) a_pass.set('fetch_list', fetch_list) a_pass.apply(self._graph) convert_pass = core.get_pass('graph_to_program_pass') desc = core.ProgramDesc() convert_pass.set_not_owned('program', desc) convert_pass.apply(self._graph) program = framework.Program._construct_from_desc(desc) if hasattr(self._program, 'lr_sheduler'): # how to share var between two different block ? lr_var_name = self._program.lr_sheduler._var_name program.lr_sheduler = self._program.lr_sheduler # Program.clone will clone lr_sheduler, so i set lr_var as # lr_sheduler attribute global_block = self._program.global_block() program.lr_sheduler.lr_var = global_block.vars[lr_var_name] # with popart, we need to support batches_per_step, what means # the shape of feed_var and feed_tensor(maybe numpy array) will # mismatch, so we set need_check_feed to False. Thus we can avoid # modify logic of run. program_global_block = program.global_block() for feed_name in feed_list: feed_var = program_global_block.var(feed_name) feed_var.desc.set_need_check_feed(False) if not hasattr(program, 'org_program'): program.org_program = self._program self._ipu_strategy.need_compile = False return program