Abstract: The disclosure herein describes elastically managing the execution of workers of multi-worker workloads on accelerator devices. A first worker of a workload is executed on an accelerator device during a first time interval. A first context switch point is identified when the first worker is in a first worker state. At the identified context switch point, a first memory state of the first worker is stored in a host memory and the accelerator device is configured to a second memory state of the second worker. The second worker is executed during a second time interval and a second context switch point is identified at the end of the second time interval when the second worker is in a state that is equivalent to the first worker state. During the intervals, collective communication operations between the workers are accumulated and, at the second context switch point, the accumulated operations are performed.

Abstract: The disclosure herein describes platform-level checkpointing for deep learning (DL) jobs. The checkpointing is performed through capturing two kinds of state data: (i) GPU state (device state), and (ii) CPU state (host state). The GPU state includes GPU data (e.g., model parameters, optimizer state, etc.) that is located in the GPU and GPU context (e.g., the default stream in GPU, various handles created by the libraries such as DNN, Blas, etc.). Only a fraction of the GPU memory is copied because the checkpointing is done in a domain-aware manner. The “active” memory contains useful data like model parameters. To be able to capture the useful data, memory management is controlled to identify which parts of the memory are active. Also, to restore the destination GPU to the same context/state, a mechanism is used to capture such state-changing events on an original GPU and replayed on a destination GPU.

Abstract: Methods, apparatus, and computer-readable media are provided for generating and applying data extraction templates. In various implementations, a corpus of structured communications such as emails may be grouped into clusters based on one or more similarities between the structured communications. A set of structural paths may be identified from structured communications of a particular cluster. One or more structural paths of the set may be classified as transient wherein a count of occurrences of one or more associated segments of text across the particular cluster satisfies a criterion. One or more transient paths may be assigned a semantic data type and/or a confidentiality designation based on various signals. A data extraction template may be generated to extract, from subsequent structured communications, segments of text associated with transient (and in some cases, non-confidential) structural paths.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for receiving a plurality of electronic documents, each electronic document being associated with an identifier that is associated with a source of the electronic document, grouping electronic documents of the plurality of electronic documents into a plurality of base sub-groups based on respective sources, for each base sub-group of the plurality of base sub-groups, automatically processing electronic documents to provide one or more templates, each template mapping content to one or more markers, and storing the one or more templates in memory, each template being accessible by one or more parsers to parse content from subsequently received electronic documents.

Abstract: Methods, apparatus, and computer-readable media are provided for generating and applying data extraction templates. In various implementations, a corpus of structured communications such as emails may be grouped into clusters based on one or more similarities between the structured communications. A set of structural paths may be identified from structured communications of a particular cluster. One or more structural paths of the set may be classified as transient wherein a count of occurrences of one or more associated segments of text across the particular cluster satisfies a criterion. One or more transient paths may be assigned a semantic data type and/or a confidentiality designation based on various signals. A data extraction template may be generated to extract, from subsequent structured communications, segments of text associated with transient (and in some cases, non-confidential) structural paths.

Abstract: Techniques for detecting poor machine translations are disclosed herein. The techniques can include various diagnostics that can be performed to determine whether to adopt a potential translated key term as a translated key term. The diagnostics include a geographic location diagnostic, a reverse translation diagnostic, a local term diagnostic, and a narrow-to-broad diagnostic.