Proto commits in microsoft/tensorflow-directml

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

2020-08-29

Add long_name field to OpMetrics PiperOrigin-RevId: 329108530 Change-Id: Ifa7df2939dac238fe96f475df16eceb137d7d8b5

2020-08-24

core/tpu/kernels/BUILD file proto target refactoring PiperOrigin-RevId: 328222137 Change-Id: I1c4339867f6e887e3647f5f60c58a7cfd0885d3f

2020-08-21

Adding TpuCompilationRpcLookup class for distributed cache support. PiperOrigin-RevId: 327894710 Change-Id: Iac493500325f49b6a0cc5158b3b3962d022935dd

[XLA] Add an optional bool is_cross_program_prefetch field to kCopyStart HLOs. This change to HLO is needed to disambiguate cross-program-prefetches and other prefetches performed over the same HloValue. This CL is in preparation for supporting freeing cross-program-prefetched buffers after their last use. PiperOrigin-RevId: 327881463 Change-Id: Id2ea6cd543589a7d49c689d44a2631a96ee9ddeb

2020-08-19

Add a device field to the object graph Variable proto. This is in preparation for supporting saving variable devices by using the experimental SAVE_VARIABLE_DEVICES SaveOptions flag. PiperOrigin-RevId: 327527925 Change-Id: I7cdd7abe92ece4be4c8ed655931c992e47ef9b53

Use the original output indices when adding a component function output to RemoteMgr. PiperOrigin-RevId: 327507408 Change-Id: Ie33d8467aec3901340ac8edd8892f28811b92c2a

Add performance analysis for TPU outside compilation. PiperOrigin-RevId: 327492744 Change-Id: Ie899823ea66e68e15fbda3578acd9cf5893554cb

Use the original output indices when adding a component function output to RemoteMgr. PiperOrigin-RevId: 327363880 Change-Id: I1f1c838a4b4decbb75d07bc3a0a6feab394477a0

Merge pull request #42448 from Intel-tensorflow:amin/nchw-nhwc-grappler-part1 PiperOrigin-RevId: 327361873 Change-Id: Id2648cf7e7597038a054fecace65b151c1326dfe

Use the original output indices when adding a component function output to RemoteMgr. PiperOrigin-RevId: 327351123 Change-Id: Ic732fca58f41d0290fc650259e90b7f48c6c9f38

2020-08-18

[tf.data service] Share dataset graphs via filesystem with RPC fallback. If a `work_dir` is configured, the dispatcher will write datasets there. When a worker needs a dataset, the dispatcher will send the worker the filesystem path of the dataset. The worker will attempt to load the dataset from filesystem. If the file reading fails, the worker falls back to sending the dispatcher an RPC request for the full dataset graph. To aid in the implementation, this CL switches worker_impl to use a DataServiceDispatcherClient instead of a raw DataService::Stub, avoiding grpc boilerplate in worker_impl. PiperOrigin-RevId: 327332591 Change-Id: I8b674d3391c7b78cb17ffd431da6eaa0e24ee30a

NCHW to NHWC conversion on CPU.

[tf.data service] Add dataset_sharing_mode option. Previously, the dataset_sharing_mode was always "rpc", with entire (potentially large) dataset graphs being sent over RPC. This CL adds another option "shared_filesystem", which shares datasets by writing them to the dispatcher's work_dir, then transmitting only the filesystem path, instead of the full dataset graph. PiperOrigin-RevId: 327130518 Change-Id: I8565689de2ce35448e8944ecc39e7ba8bb053ff9

2020-08-17

[tf.data service] Add tracking of which clients are reading from a job. This CL has no immediate functionality impact. In later CLs, we will use the client tracking to manage job lifecycles. PiperOrigin-RevId: 327056083 Change-Id: I7066bd52c9f9974f833f259990185f22e08a7370

2020-08-13

Support remote multi-device functions with outputs on any remote devices. PiperOrigin-RevId: 326549047 Change-Id: Ic6a03936e7923360b05e723a5bd4a788ec57d06b

[tf.data service] Store datasets in a "datasets" directory. This doesn't immediately change any functionality, but it will later enable us to transfer datasets to workers by sharing paths to the datasets instead of passing the full dataset definition, which may be many MB in size. PiperOrigin-RevId: 326356135 Change-Id: I7d89fa1eee53aa9f4f2f04ba37fc304a056aa68e

2020-08-11

Add XStats to XEventMetadata PiperOrigin-RevId: 326121854 Change-Id: I0c00f7ea095dab412afeaa9a1213c2ee4c44235f

Allow NdarraySpec to be written in saved model. PiperOrigin-RevId: 326121293 Change-Id: I7a4351a9ab3e0381ff5616f67d0e61880f3bb649

Add device_type to TfStatsDatabase. PiperOrigin-RevId: 326113527 Change-Id: I794cd6fa3724c8189971990abf65e3180c0b82dc

Make the semantics of repository_root more consistent. PiperOrigin-RevId: 326098252 Change-Id: I036f3943b1e35aa0543906128e0d78f59ccf1735

2020-08-10

Adding total-order comparison support in proto and HloInstruction. Specifically a comparison type attribute is added to Hlo proto so that total order comparison can be explicitly specified. A comparison expander pass is added to all compilers to expand total order comparison into equivalent implementations through type conversion. PiperOrigin-RevId: 325820826 Change-Id: I7beceb2f751ddc0be7c6b7a74037e562e7580b62

Set connect timeout based on client RPC deadline. PiperOrigin-RevId: 325805342 Change-Id: I22ec18729cf027560a479ef27596fce54bc90606

2020-08-07

Calculate TensorCore utilization PiperOrigin-RevId: 325532659 Change-Id: I0450e1ec72e22d9d9c31a26d2e16e0950c13276a

[Resubmit] If an input-output pair is configured to be must-alias(off by default), they must be aliased at runtime. PiperOrigin-RevId: 325503193 Change-Id: Ida4e46531052c40eebce5f0dff4c50914cc1f3f4

2020-08-06

Added support for zero accumulators in FTRL optimizer implementation for TPU embeddings, behind an optimization parameter setting because it hurts performance. PiperOrigin-RevId: 325137678 Change-Id: Iee7331beb6f76cdf93a7ffd33f1b8cd083b624a5

2020-08-05

[XLA] Define partial replication in HloSharding For partial replication partial tiling, add a trailing dimension to tile_assignment, representing the subgroups of replication. PiperOrigin-RevId: 325123131 Change-Id: I0e86b492e286d66d329dd9719faa0d47027f89b1

If an input-output pair is configured to be must-alias(off by default), they must be aliased at runtime. PiperOrigin-RevId: 324941010 Change-Id: I80995e25ce367ca17e5f884da55874252265f487

[tf.data service] Use the journal to keep track of registered workers. As part of this change, we stop using integer worker ids, and instead use workers addresses as their identifiers. PiperOrigin-RevId: 324927652 Change-Id: If6ef5a08aac6bf32cc603108f9045887619488f1

2020-08-04

[tf.data service] Track task creation in the dispatcher journal. This way, the dispatcher will remember which tasks exist on startup, so that when workers reconnect the dispatcher will understand their tasks and allow them to continue processing their existing tasks. Without this journaling, the dispatcher won't know what task ids to assign to new tasks, since not-yet-reconnected workers could already be using task ids. Now that task state is managed by dispatcher_state, we can replace FinishJobUpdate with FinishTaskUpdate, since dispatcher_state can identify a job as finished when its last task finishes. PiperOrigin-RevId: 324909615 Change-Id: Iee5b877aac79046662231e4b9d2a01a271a71d5d

If an input-output pair is configured to be must-alias(off by default), they must be aliased at runtime. PiperOrigin-RevId: 324905361 Change-Id: Id12e9583ec25d6464f29479c48ddef37027ef61a

[tf.data service] Write dispatcher state changes to a journal. When a work directory is configured, the dispatcher will write journal entries for its state changes to a journal within the work directory. If no work directory is configured, the dispatcher uses a NoopJournalWriter, which writes nothing. We don't yet read from the journal on dispatcher start. This support will be added in the next CL. PiperOrigin-RevId: 324749061 Change-Id: I0baf10cef05a53e0fa94139d84d5cfd284550acf

2020-08-03

Added beta parameter (as in https://research.google.com/pubs/archive/41159.pdf) to FTRL implementation for TPU embeddings. PiperOrigin-RevId: 324691573 Change-Id: I6d41c7d631e034ad95e12c6b1a3c24d4482e9171

tracking device memory transfer in steps_db. PiperOrigin-RevId: 324676713 Change-Id: I7d42f3746f35458ecf2d68a00c65f900ff673449

Add a module config option to enable hlo deduplication. PiperOrigin-RevId: 324660155 Change-Id: Ic7aac0daf851bb93b4f6c24e56b20234200efdbc

Garbage collect old WorkerSession when the restarted master task create new one. PiperOrigin-RevId: 324643608 Change-Id: I10165604d7ae03b25f15a31676d90f62aa6181be

2020-07-31

[tf.data service] Update dispatcher_impl to use dispatcher_state. This CL updates dispatcher_impl to store its dataset and job state in `DispatcherState`, and only update the state by applying state updates. The CL also changes the Task and Worker classes to be simple structs, to be consistent with the Dataset and Job classes in `DispatcherState`. No external-facing functionality is changed. The refactored functionality is covered by the tests in data_service_ops_test.py PiperOrigin-RevId: 324111209 Change-Id: I6163565c10e074d5973207039ff65ba0be071d21

2020-07-30

Add TensorCore utilization to overview. PiperOrigin-RevId: 324108255 Change-Id: I2f4bfa9d98678aa047d83531d18a6ed1967b698c

Fix duplicated profile plugin logdir. PiperOrigin-RevId: 324106374 Change-Id: Ib820ac541dbe34ad08ed0a2676677c0fce7eca81

[tf.data service] Support job creation and finishing in dispatcher_state This CL also removes CreateTaskUpdate, which is not necessary since dispatcher_state doesn't need to keep track of tasks. PiperOrigin-RevId: 324070497 Change-Id: I1ffa279f788d52211944c42b20643e7af607e572

2020-07-29

Update constant folding to allow disabling compressed tensor optimization PiperOrigin-RevId: 323853391 Change-Id: I675c26909dae99333962219482aae180f7a89cd9

2020-07-28

Added multiply_linear_by_lr documentation for FTRL for TPU embeddings. PiperOrigin-RevId: 323664047 Change-Id: I20b141e1d1eab9a3b5f6d72528d38ada350bf4ec

Create a DispatcherState class for managing dispatcher state. The goal of the class is to make it easier to keep the journal in sync with the in-memory state. Currently only the RegisterDataset state transition is implemented. Further CLs will implement the rest of the state transitions enumerated in journal.proto, and we may add more state transitions in the future. PiperOrigin-RevId: 323478950 Change-Id: I53433c6b2c45d01a3aabea26dcc62dae83e0af04

2020-07-27

Removed initial accumulator value fields since they are unused. PiperOrigin-RevId: 323420206 Change-Id: Ib1fb128d20485a689042eda8b360811334a2643e

Use proto to configure tf.data service worker server. This simplifies adding new configuration properties, so that we don't need to plumb new properties through. This also gives us a single place to document all configuration options (in the .proto file). PiperOrigin-RevId: 323416829 Change-Id: I99c1a351e0e283fdcddfa2a6d8e5768ab53da21a

Updated documentation and external source code links. PiperOrigin-RevId: 323401937 Change-Id: Ice4238aa02583a364f03410ccddaedee9246d8d2

Add tf.data service journal reader and writer. This change adds reader and writer implementations for journaling dispatcher state. The implementations read and write protocol buffers representing state changes. The records are written in TFRecord format so that we can detect corruption (TFRecord takes a CRC checksum for each record). Later CLs will update the dispatcher to write to the journal during state changes and read from the journal on restart. PiperOrigin-RevId: 323398037 Change-Id: I62eaab40219992c26df5029075a5875f212f92da

2020-07-25

[Profiler] Add more metrics. PiperOrigin-RevId: 323112988 Change-Id: Ib1147c8ada8d30be11d045b3229130665c272138

2020-07-24

Use proto to configure tf.data service worker server. This simplifies adding new configuration properties, so that we don't need to plumb new properties through. This also gives us a single place to document all configuration options (in the .proto file). PiperOrigin-RevId: 322934622 Change-Id: I547740e3c9224c7b74ecf2853672ffeb226d61d1

Use proto to configure tf.data service worker server. This simplifies adding new configuration properties, so that we don't need to plumb new properties through. This also gives us a single place to document all configuration options (in the .proto file). PiperOrigin-RevId: 322924012 Change-Id: Id53b44626ee25259b469e9e4f7d150b227b30261

2020-07-23

Use proto to configure tf.data service dispatch server. This simplifies adding new configuration properties, so that we don't need to plumb new properties through. This also gives us a single place to document all configuration options (in the .proto file). A followup CL will do the same for worker configuration. PiperOrigin-RevId: 322834156 Change-Id: I72bd7fccabf124fadb6a5675b97556c6c5c3fceb

Remove an unused proto include PiperOrigin-RevId: 322812216 Change-Id: Ic6ac2a5df5cb63627801c623533cd779504cba16

2020-07-22

Add TensorCore eligibility to overview. PiperOrigin-RevId: 322669191 Change-Id: Iaf651846869a1ea3d9881a08b0b36f48defcfb3d

Plumb TF node name into xla's argument's op metadata. PiperOrigin-RevId: 322667361 Change-Id: Ifcd875d428ce92628fc13354be9d0b4829a65f67

Change the go_package name from core_protos_go_proto to for_protos_go_proto to match the new BUILD rule name. PiperOrigin-RevId: 322617692 Change-Id: I7f5da54b0a5b6141b0c0899d12a4bc025f10c837

2020-07-20

Replace instances of "blacklist" with "denylist" where possible. See Google Developer guidelines at https://developers.google.com/style/word-list#blacklist for more information. PiperOrigin-RevId: 322242499 Change-Id: I66ea33f87811bbc734f538029d806240da91130f

2020-07-16

Update "master" to "dispatch"/"dispatcher" in tf.data service terminology. Dispatcher is more descriptive and follows the guidance in https://developers.google.com/style/word-list#master PiperOrigin-RevId: 321613785 Change-Id: Iaa576d35f0581e21278101f8b31201ba737a6865

2020-07-14

Add complex<double> tensor support in TFLite Even though we do not support complex<double> op kernels on mobile, it is inevitable to support complex<double> tensors in order to enable TF complex<double> ops via flex delegate. This CL enables the complex<double> tensor type in MLIR converter only. PiperOrigin-RevId: 321072365 Change-Id: I5ecd631339b3d5e00b3d999b9f2c6102b554cea5

2020-07-13

Fix ROCM build by removing optional keyword from protobuf definition PiperOrigin-RevId: 321025792 Change-Id: I676c7672b9de7187bac9763cf8d742321bf8804f

Add `TpuCompilationRequestProto` and helper function to serialize Compile() arguments. PiperOrigin-RevId: 320980526 Change-Id: Iacdb521e36cd2245c3eac6580efb72135f6b5ae5

2020-07-08

Replace instances of "whitelist" with "allowlist" where possible. See Google Developer guidelines at https://developers.google.com/style/word-list#blacklist for more information. PiperOrigin-RevId: 320210110 Change-Id: I480d2b1c80d7d77fdd071b7642011758988f18c0

2020-07-07

[XLA] Unify aliasing types SYSTEM/USER alias distinction is not actually used, and knowing it at compile time does not bring any advantages, as we check actual aliasing at runtime in any case. PiperOrigin-RevId: 320079893 Change-Id: I726cfe9dae0256904778a3bc3e501566aa026f9f

2020-07-06

Expose SPMD/MPMD as an option to TensorFlow. XLA/TPU has two implementations of partitioning. SPMD: ideal for per-op partitioning; MPMD: required for graph partitioning (e.g., GPipe). PiperOrigin-RevId: 319872802 Change-Id: I54e2f94445293a7036c86f5d87b4febbeb5d63b2

2020-06-30

Not yet public PiperOrigin-RevId: 319040214 Change-Id: I95edff062ee32c713df97752b29e7172bfe6aec3

2020-06-29

Expose device memory information via XRT API. PiperOrigin-RevId: 318893328 Change-Id: I1bdfc8c6fcabe7b4f9a662272aa0c40a795299da

2020-06-26

Add SaveableObjects to SavedModel. When objects are loaded from the SavedModel, they don't retain their `_gather_saveables_for_checkpoint` functions, which can result in values not being loaded from the checkpoint. This CL adds a field in the SavedModel proto that stores a save and restore function for each SaveableObject in each node. When loading into Python, the SaveableObjects are restored using the functions. PiperOrigin-RevId: 318549786 Change-Id: I688c72d7658e1bca98abf373a13a0e15a7fb83e2

Add SaveableObjects to SavedModel. When objects are loaded from the SavedModel, they don't retain their `_gather_saveables_for_checkpoint` functions, which can result in values not being loaded from the checkpoint. This CL adds a field in the SavedModel proto that stores a save and restore function for each SaveableObject in each node. When loading into Python, the SaveableObjects are restored using the functions. PiperOrigin-RevId: 318525215 Change-Id: Ic4690e0f4d24da9595c44865bbd130c323f8490a

Add SaveableObjects to SavedModel. When objects are loaded from the SavedModel, they don't retain their `_gather_saveables_for_checkpoint` functions, which can result in values not being loaded from the checkpoint. This CL adds a field in the SavedModel proto that stores a save and restore function for each SaveableObject in each node. When loading into Python, the SaveableObjects are restored using the functions. PiperOrigin-RevId: 318512603 Change-Id: I9b2b773c263703e9eb8e6114c631160ff4f7d1c1

2020-06-23

Merge pull request #40675 from zhuzilin:tiled-layout-doc-fix PiperOrigin-RevId: 317946833 Change-Id: I4cd9de065fc47143d8ccdff3552cc8bf716fe0c0

2020-06-22

[Profiler] Add eager-mode related stats and advices to the Overview Page. PiperOrigin-RevId: 317742833 Change-Id: Ib78b44b673b29dd3ff0b81fa00068f64198feb30

update

[Cleanup] Remove allowed_devices of ResourceHandle since it's no longer used. PiperOrigin-RevId: 317710941 Change-Id: Ib1920c5ee25d405290f852b725d693ee5ea09766

Use the same CompositeDevice name on remote workers as the one on a client. PiperOrigin-RevId: 317702206 Change-Id: I7068efb25eb930252f89a167108ed59c69c2078f

fix broken figures

2020-06-19

Merge pull request #40596 from reedwm:auto_mp_mkl2 PiperOrigin-RevId: 317381920 Change-Id: I8e7fe93090dafeedba1e7dccfb093d16c6e5b742

Use group_id as step_id. PiperOrigin-RevId: 317353238 Change-Id: If52b2b4872c92d3f65af8f6ce1651e8c6da7dae7

2020-06-18

Add MKL supoprt to auto_mixed_precision. This extends the auto mixed precision grappler pass to support converting nodes to bfloat16 on MKL-supported CPUs. Co-authored-by: Niranjan Hasabnis <niranjan.hasabnis@intel.com>

2020-06-16

Introduces a new experimental package that: - Defines a schema for configuring delegates - Defines a C++ plugin mechanism using the schema, so that code can support configuring arbitrary delegates without a build-time dependency PiperOrigin-RevId: 316678829 Change-Id: I36ce8a6175b550d83dfe9cf1f237a04173fb8b16

2020-06-15

Introduces a new experimental package that: - Defines a schema for configuring delegates - Defines a C++ plugin mechanism using the schema, so that code can support configuring arbitrary delegates without a build-time dependency PiperOrigin-RevId: 316566081 Change-Id: I4d36b4e155dd30fbdf57d60ef4b546304c033b1a

[tfdbg2] Add tfdbg_run_id to metadata of data dumps - A data dump file set generated by tfdbg2 can contain multiple subsets when there are multiple hosts involved in the instrumented TensorFlow job (e.g., TPUs and Parameter Servers). Currently, there is no bit in those subset of files that indicates they belong to the same instrumented TF job. - This CL addresses this problem by adding a field to the metadata proto used by those files (`tfdbg_run_id`) - The DebugEventsWriter code is revised, so that this new field is written to the metadata file of the file set on the writer's construction. - Also in this CL: remove the previous 1-arg `GetDebugEventsWriter(dump_root)` that creates the writer object if it doesn't exist at the specified dump_root. Replace it with `LookUpDebugEventsWriter(dump_root)` that only looks up the writer object and returns a non-OK status if such an object hasn't been created at `dump_root`. This makes the code less error prone by keeping only the fully-explicit, 3-arg `GetDebugEventsWriter()`. PiperOrigin-RevId: 316537044 Change-Id: Id5be0b771fbf37c0fc796f1514ed858a0e6d38f0

Introduces a new experimental package that: - Defines a schema for configuring delegates - Defines a C++ plugin mechanism using the schema, so that code can support configuring arbitrary delegates without a build-time dependency PiperOrigin-RevId: 316433209 Change-Id: Id538d0d3885bba9dd4094892915e9b6b736efd7d

2020-06-12

Add a warning message for steps dropped. PiperOrigin-RevId: 316192428 Change-Id: I1f227b32b7ec53a384f4f61c2555ca1160ffac31

Surface num dropped events as a warning message. PiperOrigin-RevId: 316138752 Change-Id: I3ea9d2c17df0dfd40c23a32ce5b1b02a80c0a9ff

sort host plane by index of xline (which is already sorted by name) PiperOrigin-RevId: 316135421 Change-Id: Ie8d3999724c129326346a2b902d4b2d5308372b2

2020-06-11

[tf.data] Implementation of `tf.data.experimental.save` and `tf.data.experimental.load`. The former makes it possible to save the output of a dataset, while the latter makes it possible to load a previously saved data. Fixes: 38483 PiperOrigin-RevId: 315991164 Change-Id: I30da604fdd489902ff4771b685e413447d3e9e9d

Merge pull request #39734 from nouiz:vlog_ptxas PiperOrigin-RevId: 315938190 Change-Id: Ib5ac12f94e3eaac65e97bdfd650af95bedf3f326

2020-06-09

Update following the comments. Also update the description.

Open sourcing some TPU-related work PiperOrigin-RevId: 315431095 Change-Id: I734632c0e5723dfca37acf53bbbd2b378b04c95d

2020-06-08

Consolidate error and warning message population. 1. Add a diagnostic message which contains error, warning and info for diagnosing profiling anomalies. Replace individual error and warning fields by the diagnostic message. 2. Add unified PopulateStepDiagnositics, PopulateOverviewDiagnositics utilities in diagnostics.cc to be used by all tool converters. 3. Add unifid GenerateDiagnosticDatatable in TfStatsProcessor for all child tools. Add a similar generate_diagnostics_table for python converters. PiperOrigin-RevId: 315356051 Change-Id: Ic1f1d6d43b3fec850ccb04c3184f5ba7bbd694af

Add XLA_FLAGS=--xla_gpu_gpuasm_extra_flags=...

2020-06-04

Update naming of embedding in callback for TensorBoard integration. PiperOrigin-RevId: 314693951 Change-Id: I6a3aba5b98ce6865da8814eda554bac1fb6bcaea

2020-06-03

Implement GPU device priority for virtual devices. - This adds a field in virtual device options to specify priority for virtual devices. - When the priority is specified, it will be used to create the cuda streams for the virtual device with the given priority. - This is backwards compatible with no priorities specified. When no priorities specified, the current implementation of creating a stream without any priority will continue while any non-zero priorities specified will be used to create streams with that priority. PiperOrigin-RevId: 314470276 Change-Id: I4943f71e901245fb21b6f7e833adbdcd8126f1fa

2020-06-02

[Profiler] Explicitly stores the input percentage in the input-analysis proto. PiperOrigin-RevId: 314408625 Change-Id: I2bd23624a45949820543b3681534216fb20a7c3d

2020-05-29

Internal change PiperOrigin-RevId: 313850352 Change-Id: I89584b0bcb4409eb74d21e31fb0eb68844186707

Implement GPU device priority for virtual devices. - This adds a field in virtual device options to specify priority for virtual devices. - When the priority is specified, it will be used to create the cuda streams for the virtual device with the given priority. - This is backwards compatible with no priorities specified. When no priorities specified, the current implementation of creating a stream without any priority will continue while any non-zero priorities specified will be used to create streams with that priority. PiperOrigin-RevId: 313814564 Change-Id: Ie580ea3f8c7fc7248e40e1ae4c898f0a029fba35

2020-05-27

Remove the xla_gpu_unsafe_fallback_to_driver_on_ptxas_error flag. PiperOrigin-RevId: 313389132 Change-Id: Ic97116d9b471e96822ee28032ce0ddef5616a4f0

2020-05-26

Update server_lib API to be more consistent with tf.distribute.Server. This is in preparation for exporting MasterServer and WorkerServer in the public API. This CL also addresses an issue where we relied on counting the number of tasks in job 0 to determine when a worker has registered with the master. Now we directly query the master for how many workers are registered. PiperOrigin-RevId: 313269683 Change-Id: Ie4284a8ca91bd87fd366761d055761384654aef3

2020-05-21

Add memory cost breakdown per operation type and memory space to OpMetrics. PiperOrigin-RevId: 312741555 Change-Id: Id6666a8c6b9d67fe443154fb135cadeeeaecebdc

2020-05-20

[Profiler] Add optimization advice for reducing the host-to-TPU data transfer time. PiperOrigin-RevId: 312388184 Change-Id: I2fc8a60af6724467e447026dde7a8d6925ed1357

2020-05-19

Move compression_utils to core/data. This is in preparation for adding a CompressElementOp, which will use CompressElement to compress a dataset element in a tf.data service agnostic way. PiperOrigin-RevId: 312197651 Change-Id: I3558b2f5036dcf4c91ed9059a7b896351c79da40

2020-05-16

Adding an option to tensor tracer to create a suffix folder based on the fingerprint of the tf.graph. If use_fingerprint_subdirectory is provided, then the TensorTracer summaries will be written under <trace_dir>/<fingerprint>. If there are changes to the graph, the changes will be listed under different fingerprints. PiperOrigin-RevId: 311834837 Change-Id: I9dfbabfeb7fbe58a2a47c2581474ed86647781dc