def _allreduce_cond(tensor, *args, **kwargs): def allreduce_fn(): return allreduce(tensor, *args, **kwargs) def id_fn(): return tensor return tf.cond(size_op() > 1, allreduce_fn, id_fn)
def _grouped_allreduce_cond(tensors, *args, **kwargs): def allreduce_fn(): return grouped_allreduce(tensors, *args, **kwargs) def id_fn(): return tensors return tf.cond((size_op() > 1) if int(os.environ.get("HOROVOD_ELASTIC", 0)) else tf.convert_to_tensor(size() > 1), allreduce_fn, id_fn)
def _grouped_allreduce_cond(tensors, *args, process_set=global_process_set, **kwargs): def allreduce_fn(): return grouped_allreduce(tensors, *args, process_set=process_set, **kwargs) def id_fn(): return tensors return tf.cond(tf.logical_and( tf.equal(process_set_included_op(process_set.process_set_id), 1), tf.greater(size_op(process_set.process_set_id), 1)) if int(os.environ.get("HOROVOD_ELASTIC", 0)) else ( tf.convert_to_tensor(process_set.included() and process_set.size() > 1)), allreduce_fn, id_fn)
def reducescatter(tensor, device_dense='', compression=Compression.none, op=Average, name=None, process_set=global_process_set): """Perform a reducescatter on a tf.Tensor. This function performs a bandwidth-optimal reduce and scatter on the input tensor. Arguments: tensor: tf.Tensor or tf.Variable to reduce. The shape of the input must be identical across all ranks. device_dense: Device to be used for dense tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_REDUCESCATTER. compression: Compression algorithm used to reduce the amount of data sent and received by each worker node. Defaults to not using compression. op: The reduction operation to combine tensors across different ranks. Defaults to Average. process_set: Process set object to limit this operation to a subset of Horovod processes. Default is the global process set. name: A name of the reduce_scatter operation Returns: A tensor of the same rank and type as `tensor`, summed across all processes. The shape is identical to the input shape, except for the first dimension, which will be divided across the different Horovod processes. """ # Averaging happens in framework code, so translate that to Sum for the actual call true_op = Sum if op == Average else op with tf.device(device_dense): horovod_size = tf.cast( size_op(process_set_id=process_set.process_set_id) if int( os.environ.get("HOROVOD_ELASTIC", 0)) else process_set.size(), dtype=tensor.dtype) tensor_compressed, ctx = compression.compress(tensor) reduced_tensor_compressed = _reducescatter(tensor_compressed, op=true_op, name=name, process_set=process_set) reduced_tensor = compression.decompress(reduced_tensor_compressed, ctx) new_tensor = (reduced_tensor / horovod_size) if op == Average else reduced_tensor return new_tensor
def allreduce(tensor, average=None, device_dense='', device_sparse='', compression=Compression.none, op=None, prescale_factor=1.0, postscale_factor=1.0, name=None): """Perform an allreduce on a tf.Tensor or tf.IndexedSlices. This function performs a bandwidth-optimal ring allreduce on the input tensor. If the input is an tf.IndexedSlices, the function instead does an allgather on the values and the indices, effectively doing an allreduce on the represented tensor. Arguments: tensor: tf.Tensor, tf.Variable, or tf.IndexedSlices to reduce. The shape of the input must be identical across all ranks. average: .. warning:: .. deprecated:: 0.19.0 Use `op` instead. Will be removed in v0.21.0. device_dense: Device to be used for dense tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. device_sparse: Device to be used for sparse tensors. Uses GPU by default if Horovod was built with HOROVOD_GPU_OPERATIONS. compression: Compression algorithm used to reduce the amount of data sent and received by each worker node. Defaults to not using compression. op: The reduction operation to combine tensors across different ranks. Defaults to Average if None is given. prescale_factor: Multiplicative factor to scale tensor before allreduce. postscale_factor: Multiplicative factor to scale tensor after allreduce. name: A name of the allreduce operation Returns: A tensor of the same shape and type as `tensor`, summed across all processes. """ op = handle_average_backwards_compatibility(op, average) if isinstance(tensor, tf.IndexedSlices): # TODO: Need to fix this to actuall call Adasum if op == Adasum: raise NotImplementedError( 'The Adasum reduction does not currently support sparse tensors. As a ' 'workaround please pass sparse_as_dense=True to DistributedOptimizer' ) with tf.device(device_sparse): # For IndexedSlices, do two allgathers instead of an allreduce. horovod_size = tf.cast(size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.values.dtype) values = allgather(tensor.values) indices = allgather(tensor.indices) # To make this operation into an average, divide allgathered values by # the Horovod size. new_values = (values / horovod_size) if op == Average else values return tf.IndexedSlices(new_values, indices, dense_shape=tensor.dense_shape) else: average_in_framework = False if rocm_built(): # For ROCm, perform averaging at framework level average_in_framework = op == Average or op == Adasum op = Sum if op == Average else op with tf.device(device_dense): horovod_size = tf.cast(size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.dtype) tensor_compressed, ctx = compression.compress(tensor) summed_tensor_compressed = _allreduce( tensor_compressed, op=op, prescale_factor=prescale_factor, postscale_factor=postscale_factor, name=name) summed_tensor = compression.decompress(summed_tensor_compressed, ctx) if op == Adasum: if 'CPU' not in tensor.device and gpu_available('tensorflow'): if nccl_built(): if not is_homogeneous: raise NotImplementedError( 'Running GPU Adasum on heterogeneous cluster is not supported yet.' ) elif not check_num_rank_power_of_2( int(size() / local_size())): raise NotImplementedError( 'Running GPU Adasum with non-power of 2 nodes is not supported yet.' ) if rocm_built(): horovod_local_size = tf.cast( local_size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else local_size(), dtype=tensor.dtype) new_tensor = summed_tensor / horovod_local_size else: new_tensor = summed_tensor else: warnings.warn( 'Adasum reduction does not currently support GPU reduction using MPI. Tensors ' 'are copied to CPU memory instead. To use Adasum for GPU reduction, please ' 'compile Horovod with HOROVOD_GPU_OPERATIONS=NCCL.' ) new_tensor = summed_tensor else: if not check_num_rank_power_of_2(size()): raise NotImplementedError( 'Running Adasum with non-power of 2 ranks is not supported yet.' ) new_tensor = summed_tensor else: if rocm_built(): new_tensor = (summed_tensor / horovod_size ) if average_in_framework else summed_tensor else: new_tensor = summed_tensor return new_tensor
def grouped_allreduce(tensors, average=None, device_dense='', device_sparse='', compression=Compression.none, op=None, prescale_factor=1.0, postscale_factor=1.0): if not tensors: return tensors op = handle_average_backwards_compatibility(op, average) average_in_framework = False if rocm_built(): # For ROCm, perform averaging at framework level average_in_framework = op == Average or op == Adasum op = Sum if op == Average else op if any(isinstance(t, tf.IndexedSlices) for t in tensors): # TODO: Need to fix this to actuall call Adasum if op == Adasum: raise NotImplementedError( 'The Adasum reduction does not currently support sparse tensors. As a ' 'workaround please pass sparse_as_dense=True to DistributedOptimizer' ) with tf.device(device_sparse): new_values = [] for tensor in tensors: # For IndexedSlices, do two allgathers instead of an allreduce. horovod_size = tf.cast(size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.values.dtype) values = allgather(tensor.values) indices = allgather(tensor.indices) # To make this operation into an average, divide allgathered values by # the Horovod size. new_values += (values / horovod_size) if op == Average else values return [ tf.IndexedSlices(x, indices, dense_shape=t.dense_shape) for x, t in zip(new_values, tensors) ] else: with tf.device(device_dense): tensors_compressed, ctxs = zip( *[compression.compress(tensor) for tensor in tensors]) summed_tensors_compressed = _grouped_allreduce( tensors_compressed, op=op, prescale_factor=prescale_factor, postscale_factor=postscale_factor) summed_tensors = [ compression.decompress(t, ctx) for t, ctx in zip(summed_tensors_compressed, ctxs) ] if op == Adasum: if 'CPU' not in tensor.device and gpu_available('tensorflow'): if nccl_built(): if not is_homogeneous: raise NotImplementedError( 'Running GPU Adasum on heterogeneous cluster is not supported yet.' ) elif not check_num_rank_power_of_2( int(size() / local_size())): raise NotImplementedError( 'Running GPU Adasum with non-power of 2 nodes is not supported yet.' ) if rocm_built(): new_tensors = [] for tensor in summed_tensors: horovod_local_size = tf.cast( local_size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else local_size(), dtype=tensor.dtype) new_tensors += tensor / horovod_local_size else: new_tensors = summed_tensors else: warnings.warn( 'Adasum reduction does not currently support GPU reduction using MPI. Tensors ' 'are copied to CPU memory instead. To use Adasum for GPU reduction, please ' 'compile Horovod with HOROVOD_GPU_OPERATIONS=NCCL.' ) new_tensors = summed_tensors else: if not check_num_rank_power_of_2(size()): raise NotImplementedError( 'Running Adasum with non-power of 2 ranks is not supported yet.' ) new_tensors = summed_tensors else: if rocm_built(): new_tensors = [] for tensor in summed_tensors: horovod_size = tf.cast(size_op() if int( os.environ.get("HOROVOD_ELASTIC", 0)) else size(), dtype=tensor.dtype) new_tensors += ( tensor / horovod_size) if average_in_framework else tensor else: new_tensors = summed_tensors return new_tensors