Abstract: Methods, systems, apparatuses, and computer-readable media are disclosed for identifying optimization opportunities for subgraphs representing programs that perform operations on data. Tests may be performed on the subgraph to identify opportunities for fusing one or more nodes, such as more data nodes and/or one or more operations nodes, of the subgraph. For instance, optimization through node fusion may be possible where the subgraph comprises all nodes and all edges in a path between two operations nodes in the subgraph. Based on such identified optimization opportunities, steps such as node fusion may be performed, and these operations may allow for the program to be run in a manner that preserves computing resources without introducing data processing errors.

Abstract: A method for managing an intermediate representation from a program is executed by one or more processors, and includes extracting, from the program, information on data for input and output and information on operation, generating an intermediate representation from the program using the extracted information on data and the extracted information on operation, storing, in a database, a corresponding relationship between the program and the intermediate representation, storing execution information on operation of the intermediate representation, and deleting at least a part of the intermediate representation based on the execution information.

Abstract: A method for compilation optimization using activation recalculation is provided, which is performed by one or more processors, and includes receiving a source program, determining an operation to be executed in a processor based on the source program, and determining whether or not the operation to be executed corresponds to the operation for activation recalculation, thereby automatically classifying the operation to be executed into a first operation type corresponding to the operation for activation recalculation or a second operation type not corresponding to the operation for activation recalculation.

Abstract: A method for compiling an application is executed by one or more processors, and includes acquiring profiling information of a system on which an application is to be executed, generating a cost model based on the profiling information, acquiring an intermediate representation of at least a portion of the application, applying compiler passes to the intermediate representation and generating a compiled graph, and using the cost model, determining an expected execution time for the compiled graph.

Abstract: A method for pipeline parallelism compiling is provided, which is executed by one or more processors, and includes receiving a source program associated with training of a machine learning model, determining, based on the source program, a plurality of operation groups including operations executed on training data of the machine learning model, generating a plurality of micro-batches from the training data, and determining, for each of the plurality of micro-batches, a plurality of operation sets corresponding to the plurality of operation groups.

Abstract: A method for performing checkpointing for fault tolerance of an application is provided, which is performed by one or more processors, and includes storing accelerator data associated with an application at a specific time point as a checkpoint, storing operation data associated with the application performed after the specific time point, and performing application fault tolerance based on the checkpoint and the stored operation data.

Abstract: A method for generating an intermediate representation for a program for execution on an accelerator is executed by one or more processors, and includes hooking information on instruction from a program, determining whether the hooked information on instruction is associated with an accelerator, if it is determined that the information on instruction is associated with the accelerator, generating a first intermediate representation for the instruction using information on input and output data and information on instruction included in the instruction, and generating a second intermediate representation for the program for one or more accelerators using the first intermediate representation, and the first intermediate representation and the second intermediate representation include a plurality of data nodes, one or more operation nodes, and a plurality of edges indicating an input and output relationship between the plurality of data nodes and the one or more operation nodes.

Abstract: Provided is a fault tolerance method which is performed by one or more processors, and which includes receiving an application execute command, executing a main process of an application in response to the execute command, receiving, by a split execution module, information on a plurality of devices associated with the execution of the application from an orchestrator, executing, by the split execution module, a sub-process for each of the plurality of devices using the information on the plurality of devices, and performing, by the split execution module, fault tolerance associated with the execution of the application using an idle device, if a failure occurs in at least some of the plurality of devices.

Abstract: Embodiments disclosed herein relate to a method for GPU memory management that observes the deep learning of a deep neural network performed by a GPU and reduces the amount of GPU memory used, thereby overcoming limitations attributable to the memory size of the GPU and allowing the more effective performance of the deep learning, and a computing device for performing the same. According to an embodiment, there is disclosed a method for GPU memory management for a deep neural network, the method being performed by a computing device including a GPU and a CPU, the method including: generating a schedule for GPU memory management based on the processing of a unit operation, included in the deep neural network, by the GPU; and moving data required for deep learning of the deep neural network between GPU memory and CPU memory based on the schedule.

Abstract: The present disclosure relates to a method for generating a program for use in an accelerator for deep learning. The method may include receiving, by a computing device, a deep learning application, generating an element-wise operation list included in the deep learning application, generating an intermediate expression from the element-wise operation list, and generating, based on the intermediate expression, a program for use in an accelerator for the deep learning application.

Abstract: A method for automatically optimizing a program based on reinforcement learning includes (a) receiving an input for a source program, which includes a fixed parameter and variable parameter, (b) generating the source program based on the received input, (c) converting the source program into an object program, (d) executing the converted object program to measure a performance of the executed object program, (e) inputting the variable parameter and the measured performance into a machine learning model, and outputting a variation of the variable parameter, and (f) regenerating a source program reflecting the variation of the variable parameter.