These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)
| Commit: | 716954e | |
|---|---|---|
| Author: | Antonio Sanchez | |
| Committer: | Copybara-Service | |
Add missing license headers. PiperOrigin-RevId: 934549510
| Commit: | 9a4cd70 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Support multipass profiling in multihost_hlo_runner. PiperOrigin-RevId: 919831591
| Commit: | 79cf295 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Default behaviour correction for GetSnapshotRequest PiperOrigin-RevId: 931213870
The documentation is generated from this commit.
| Commit: | abc490c | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add snapshot_session_id to GetSnapshotRequest proto. PiperOrigin-RevId: 930926325
| Commit: | 6ad7fde | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Implement snapshot_session_id validation and override in GetSnapshot. PiperOrigin-RevId: 930605458
| Commit: | 2dea0f9 | |
|---|---|---|
| Author: | Subham Soni | |
| Committer: | Copybara-Service | |
Standardize XProf hostname resolution logic PiperOrigin-RevId: 898840873
| Commit: | f26e18f | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add perf counter sampling options. PiperOrigin-RevId: 904070885
| Commit: | 5860113 | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add perf_counters option to ProfilerOptions PiperOrigin-RevId: 904434657
| Commit: | 584a493 | |
|---|---|---|
| Author: | Ilya Tikhonovskiy | |
| Committer: | Copybara-Service | |
Add aggregation mode to cupti collector. The goal of the pr is to generate a small, lightweight aggregated profile that could be collected frequently and would not slow down the execution. PiperOrigin-RevId: 885476831
| Commit: | b7d698f | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Creating UI options for trace mode when capturing profiles using sampling mode PiperOrigin-RevId: 812867750
| Commit: | bf9f65b | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add StopContinuousProfiling RPC to ProfilerService. PiperOrigin-RevId: 862130665
| Commit: | b10432d | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add JAX version metadata to xspace PiperOrigin-RevId: 840138208
| Commit: | e109aca | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add stop_continuous_profiling to pywrap profiler plugin. PiperOrigin-RevId: 859542515
| Commit: | 1c307f5 | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add RPC to stop continuous profiling. PiperOrigin-RevId: 859516646
| Commit: | 7415bdc | |
|---|---|---|
| Author: | Subham Soni | |
| Committer: | Copybara-Service | |
Add override_hostname to profiler_options This will allow a follow-up PR that allows utilizing this proto. PiperOrigin-RevId: 853576203
| Commit: | 3cb4f30 | |
|---|---|---|
| Author: | Subham Soni | |
| Committer: | Copybara-Service | |
Use override_hostname from ProfileOptions PiperOrigin-RevId: 853561309
| Commit: | 1450399 | |
|---|---|---|
| Author: | Aditya Sharma | |
| Committer: | Copybara-Service | |
Add ContinuousProfiling and GetSnapshot RPCs to the profiler service. PiperOrigin-RevId: 851180780
| Commit: | a60ffab | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Enable circular buffer tracing for TPU PiperOrigin-RevId: 848119258
| Commit: | 9182be0 | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add Python bindings for continuous profiling and snapshot retrieval. PiperOrigin-RevId: 843151720
| Commit: | 11676f9 | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Add ContinuousProfiling and GetSnapshot RPCs to the profiler service. PiperOrigin-RevId: 843143959
| Commit: | a765cdd | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
The comment for XEvents within an XLine is updated to specify that partial overlap is not allowed, while nesting is still permitted. PiperOrigin-RevId: 838916383
| Commit: | a8fece0 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add int32 type to AdvancedConfigValue in ProfilerOptions. This allows `AdvancedConfigValue` to store `int32` values, expanding the types of configurations that can be represented. PiperOrigin-RevId: 830774830
| Commit: | b2a69ed | |
|---|---|---|
| Author: | Matt Hurd | |
| Committer: | Copybara-Service | |
Add int32 type to AdvancedConfigValue in ProfilerOptions. This allows `AdvancedConfigValue` to store `int32` values, expanding the types of configurations that can be represented. PiperOrigin-RevId: 830766882
| Commit: | a1cb380 | |
|---|---|---|
| Author: | Matt Hurd | |
| Committer: | Copybara-Service | |
[xprof] Add support for tracemark_lower and tracemark_upper profiler options PiperOrigin-RevId: 816343394
| Commit: | 8a644da | |
|---|---|---|
| Author: | Matt Hurd | |
| Committer: | Copybara-Service | |
Add session_id to profiler_options This will allow a follow-up PR that allows utilizing this proto. PiperOrigin-RevId: 824709715
| Commit: | e71dfd9 | |
|---|---|---|
| Author: | Bryan Massoth | |
| Committer: | Copybara-Service | |
Add a profiler for subprocesses. PiperOrigin-RevId: 805455652
| Commit: | 7c4f81f | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Set remaining profile options and perform validation on the advanced_configuration keys. PiperOrigin-RevId: 761811794
| Commit: | d71df2f | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
[AutoPGLE] Prevent an AutoPGLE to run if user launched an external profiler. PiperOrigin-RevId: 740804528
| Commit: | 961956b | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
[AutoPGLE] Prevent an AutoPGLE to run if user launched an external profiler. PiperOrigin-RevId: 739431800
| Commit: | 5f4fcc6 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
[AutoPGLE] Prevent an AutoPGLE to run if user launched an external profiler. PiperOrigin-RevId: 739109278
| Commit: | 8bc89f8 | |
|---|---|---|
| Author: | Sannidhya Chauhan | |
| Committer: | Copybara-Service | |
Introduce an advanced configuration option for the profiler. PiperOrigin-RevId: 738702030
| Commit: | e3ead3e | |
|---|---|---|
| Author: | Julia Guo | |
| Committer: | Copybara-Service | |
Move memory_profile.proto to OSS to prepare for OSS benchmarking PiperOrigin-RevId: 736560797
| Commit: | add1173 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
This CL updates HostTracer to use Traceme Filter mask from profiling request. Also add utility functions to prepare filter mask. PiperOrigin-RevId: 700753727
| Commit: | 5f3ef51 | |
|---|---|---|
| Author: | David Dunleavy | |
| Committer: | Copybara-Service | |
Move `tsl/protobuf/error_codes.proto` to `xla/tsl/protobuf` PiperOrigin-RevId: 682485437
| Commit: | 7804cf3 | |
|---|---|---|
| Author: | David Dunleavy | |
| Committer: | Copybara-Service | |
Move `tsl/protobuf/*` besides `error_codes.proto` to `xla/tsl/protobuf` `error_codes.proto` will be moved in a separate change PiperOrigin-RevId: 678848832
| Commit: | 4f6d2cd | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Changes to the test log proto for TSL changes PiperOrigin-RevId: 668092486
| Commit: | 71c704c | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Proto bfc_memory_map change to compiler/xla/tsl PiperOrigin-RevId: 665044485
| Commit: | e213475 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add `ReportInfoToService` to coordination service for runtime info/error reporting and aggregation. This allows the coordinator to become a central place to gather runtime information and report it. PiperOrigin-RevId: 660545522
| Commit: | a78f990 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add long polling as a new way to propagate error in coordination service. PiperOrigin-RevId: 656228596
| Commit: | 4abf8c8 | |
|---|---|---|
| Author: | David Dunleavy | |
| Committer: | Copybara-Service | |
Move `tensorflow/tsl/protobuf` to `xla/tsl/protobuf` PiperOrigin-RevId: 644440378
| Commit: | ccf06a2 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Allow optional KV store overwrites. Cleanups: use std::string_view, migrate away from TF_ASSERT/EXPECT. PiperOrigin-RevId: 639131697
| Commit: | ff3217b | |
|---|---|---|
| Author: | Shanbin Ke | |
| Committer: | Copybara-Service | |
PR #11449: [XLA:GPU] replace is_causal_mask with enum mask_type to support more cudnn mask generation Imported from GitHub PR https://github.com/openxla/xla/pull/11449 * replace is_causal_mask in backend_config with enum mask_type. * cudnn also supports padding mask and alibi mask. Copybara import of the project: -- 5780ffd6622978c0a173d90cc562d32f354bfbfd by cjkkkk <ske@nvidia.com>: replace is_causal_mask with mask_type to support more mask -- 73555dba8f8cc1319cf75ed9cf8fe49ad88948bd by cjkkkk <ske@nvidia.com>: add newline -- 2fed993035a2e5c17c990b60b7268ea9dbaeb7e5 by cjkkkk <ske@nvidia.com>: add is_causal_mask back in backend_config Merging this change closes #11449 PiperOrigin-RevId: 625595798
| Commit: | 00b7447 | |
|---|---|---|
| Author: | David Dunleavy | |
| Committer: | Copybara-Service | |
Move tsl/distributed_runtime to xla/tsl/distributed_runtime PiperOrigin-RevId: 623975602
| Commit: | 8bd84de | |
|---|---|---|
| Author: | Philipp Hack | |
| Committer: | Copybara-Service | |
PR #8225: Layer Norm Gradient Fusion Imported from GitHub PR https://github.com/openxla/xla/pull/8225 Rewrites the backward graph of layer norm patterns (X, S, E, N, DY) -> (DX, DS, DB), where X, S, E, N and DY are the input to the forward layer norm, scale, expectation, norm factor and gradient w.r.t. the result of the forward layer norm, and DX, DS and DB are the gradients w.r.t. the input, scale and bias tensors, into a Custom Call to the cuDNN library. Copybara import of the project: -- 657202adda82f6fc60a88624ce2fb1c152bb5ed6 by Philipp Hack <phack@nvidia.com>: Support for layer norm gradient fusion. -- 91434d8054bbe15c2a94f864bdda3ebc10437805 by Philipp Hack <phack@nvidia.com>: Support for layer norm gradient fusion. -- 816d4c770e797d1558897a30ed033707ac6b0123 by Philipp Hack <phack@nvidia.com>: Support for layer norm gradient fusion. -- 946b0c8d9b3f018318aff846b888a453e1948612 by Philipp Hack <phack@nvidia.com>: Support for layer norm gradient fusion. Merging this change closes #8225 PiperOrigin-RevId: 607521087
| Commit: | 17ce560 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Adds compressed_string type to XStat PiperOrigin-RevId: 576496122
| Commit: | bbee6bd | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Added placeholder for internal RPC option. PiperOrigin-RevId: 565319212
| Commit: | 9ca4830 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Add escape hatch for users who do not want to enable coordination service by default. PiperOrigin-RevId: 560849974
| Commit: | 8606673 | |
|---|---|---|
| Author: | Philipp Hack | |
| Committer: | Copybara-Service | |
PR #60807: FP8 Convolutions in XLA Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/60807 Enables scaled convolutions of the form (X, W, x_scale, w_scale, y_scale) -> Y, where the input X, the filter W and the output Y are based on the `F8E4M3FN` and `F8E5M2` data types and x_scale, w_scale and y_scale are their scaling factors. Copybara import of the project: -- 8a30aa731c21612fe098a6b620a54922578611c2 by Philipp Hack <phack@nvidia.com>: Support for FP8 convolutions in XLA. -- caade6453519ad2531ebcf8f206e40187a1687ca by Philipp Hack <phack@nvidia.com>: Support for FP8 convolutions in XLA. -- ecd080bd6c64682f6bee62f4455ea2c37c279f26 by Philipp Hack <phack@nvidia.com>: Support for FP8 convolutions in XLA. -- da22a881a3d24fd4f357207034ba6c596aa414d0 by Philipp Hack <phack@nvidia.com>: Support for FP8 convolutions in XLA. Merging this change closes #60807 PiperOrigin-RevId: 551973730
| Commit: | 2862c3d | |
|---|---|---|
| Author: | TJ Xu | |
| Committer: | Copybara-Service | |
PR #3886: [NVIDIA XLA:GPU] Introducing training support for cudnn fused mha (stream executor) Imported from GitHub PR https://github.com/openxla/xla/pull/3886 This is the commit to include only stream executor changes. Rewriter changes are in this [pr](https://github.com/openxla/xla/pull/3726) lhlo lowering logic is in this [pr](https://github.com/openxla/xla/pull/3886) Copybara import of the project: -- 0541a4c40a3cec2bc9b348043a02713602a28793 by TJ <tjx@nvidia.com>: Introducing training support for cudnn fused mha This is the commit to include only stream executor changes -- abec4e1134e4d934fed9231f6e29833a687e2d58 by TJ <tjx@nvidia.com>: Replace tsl::error with absl:: calls -- 4af3ef74ca208348cb5a90a3d8f000e4e2d1c1d8 by TJ <tjx@nvidia.com>: Place some of the vlog statements before std::move -- 0606d44ed2ea85a7a4c80dfb68237ffcc51a43cc by TJ <tjx@nvidia.com>: Address clang-tidy issues Merging this change closes #3886 PiperOrigin-RevId: 546371252
| Commit: | 3cbd099 | |
|---|---|---|
| Author: | Yin Zhang | |
| Committer: | Copybara-Service | |
1 line change: Fix comment typo in xplane.proto PiperOrigin-RevId: 544699109
| Commit: | c3ef667 | |
|---|---|---|
| Author: | George Karpenkov | |
| Committer: | Copybara-Service | |
Move autotuning protos to XLA The protobufs are mostly used and modified inside of XLA, which also provide most of the GPU performance. PiperOrigin-RevId: 544346444
| Commit: | 7d8807d | |
|---|---|---|
| Author: | Jake Hall | |
| Committer: | Copybara-Service | |
PR #3200: Add support for float8_e4m3fnuz and float8_e5m2fnuz. Imported from GitHub PR https://github.com/openxla/xla/pull/3200 This adds support for the two FP8 types `float8_e4m3fnuz` and `float8_e5m2fnuz` to XLA similar to `float8_e4m3fn`, `float8_e4m3b11`, and `float8_e5m2`. Copybara import of the project: -- 3b96f8fe219c1ac1bec5c4b99ff9c51148706981 by Jake Hall <jakeh@graphcore.ai>: Add support for float8_e4m3fnuz and float8_e5m2fnuz. Merging this change closes #3200 PiperOrigin-RevId: 544198797
| Commit: | df7e113 | |
|---|---|---|
| Author: | Tao Wang | |
| Committer: | Copybara-Service | |
Move ProfiledInstructionsProto from `tensorflow/compiler/xla/...` to `tensorflow/tsl/profiler/protobuf/...` PiperOrigin-RevId: 542757964
| Commit: | 672204a | |
|---|---|---|
| Author: | David Majnemer | |
| Committer: | Copybara-Service | |
[XLA] Add support for E4M3B11 for CPU & GPU PiperOrigin-RevId: 528942681
| Commit: | 90a23e5 | |
|---|---|---|
| Author: | Jie Sun | |
| Committer: | Copybara-Service | |
change Trace/TraceEvent etc to tsl::profiler package. PiperOrigin-RevId: 522821950
| Commit: | 22ad2ad | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
[Coord Service] Guard silent reconnect upon restart feature with a config field. Disable this by default so that unexpected reconnects upon restart will return an error. Context: this feature partly caused restart loops under certain conditions. PiperOrigin-RevId: 520488973
| Commit: | 6bbc6b8 | |
|---|---|---|
| Author: | Ayan Moitra | |
| Committer: | Copybara-Service | |
PR #60145: [NVIDIA XLA:GPU] Fused MHA Support in XLA GPU: SE Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/60145 cuDNN currently supports the following patterns for Multi-headed attention: BMM1 - BMM2 BMM1 - Scale - Bias - Mask - Softmax - BMM2 BMM1 - Scale - Bias - Mask - Softmax - Dropout - BMM2 BMM1 - Scale - Mask - Softmax - BMM2 BMM1 - Scale - Mask - Softmax - Dropout - BMM2 BMM1 - Softmax - Dropout - BMM2 This PR adds support for the stream executor for these patterns. Copybara import of the project: -- d65ed4bb5de2607f09eeb5d065caa3612d750a85 by Ayan Moitra <amoitra@nvidia.com>: Revert "Revert: PR #60047: Fused MHA Support in XLA GPU: Stream Executor changes" This reverts commit 7a4d2e8b1a503a45ff86d40b0387fc832302379f. -- 532ed5cf024a72716ed1040d973d74eb5be471d3 by Ayan Moitra <amoitra@nvidia.com>: Fused MHA Support in XLA GPU: SE Merging this change closes #60145 PiperOrigin-RevId: 520458719
| Commit: | e31f005 | |
|---|---|---|
| Author: | Peter Hawkins | |
| Committer: | Copybara-Service | |
Revert: PR #60047: Fused MHA Support in XLA GPU: Stream Executor changes Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/60047 cuDNN currently supports the following patterns for Multi-headed attention: BMM1 - BMM2 BMM1 - Scale - Bias - Mask - Softmax - BMM2 BMM1 - Scale - Bias - Mask - Softmax - Dropout - BMM2 BMM1 - Scale - Mask - Softmax - BMM2 BMM1 - Scale - Mask - Softmax - Dropout - BMM2 BMM1 - Softmax - Dropout - BMM2 This PR adds support for the stream exec... PiperOrigin-RevId: 519770880
| Commit: | 09c41e9 | |
|---|---|---|
| Author: | Ayan Moitra | |
| Committer: | Copybara-Service | |
PR #60047: Fused MHA Support in XLA GPU: Stream Executor changes Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/60047 cuDNN currently supports the following patterns for Multi-headed attention: BMM1 - BMM2 BMM1 - Scale - Bias - Mask - Softmax - BMM2 BMM1 - Scale - Bias - Mask - Softmax - Dropout - BMM2 BMM1 - Scale - Mask - Softmax - BMM2 BMM1 - Scale - Mask - Softmax - Dropout - BMM2 BMM1 - Softmax - Dropout - BMM2 This PR adds support for the stream executor for these patterns. Copybara import of the project: -- c957cb879b9757b5c524106aab066a09dab32dfd by Ayan Moitra <amoitra@nvidia.com>: Fused MHA Support in XLA GPU: Stream Executor changes -- 68785164a1ba14457148f1a7ed616183dec7c481 by Ayan Moitra <amoitra@nvidia.com>: Test build failure fix -- 99292e9540c02708b9acec872a60140000c31a4e by Ayan Moitra <amoitra@nvidia.com>: Possible fix for build failure -- 7b47ff4c5b7c429285ca437a308dd81ea66c1fd1 by Ayan Moitra <amoitra@nvidia.com>: Still fixing build test build fail-cannot repro locally -- d86e51b1e4a4a57ed50efc019f2f73d03d325cfa by Ayan Moitra <amoitra@nvidia.com>: continue fixing build test build fail-cannot repro locally -- 0bf1352234fddb20ddfec6fdc01b2806fb8c820b by Ayan Moitra <amoitra@nvidia.com>: try fix build issue Merging this change closes #60047 PiperOrigin-RevId: 519675175
| Commit: | 1a0807a | |
|---|---|---|
| Author: | Yang Chen | |
| Committer: | Copybara-Service | |
#tf-data-service Improve error handling for SnapshotManager. If the snapshot manager receives an error from a worker: 1. It writes a StatusProto to an ERROR file. The error status can be recovered if the dispatcher restarts. 2. It returns empty task lists to other workers. The workers will then cancel the ongoing work here: https://github.com/tensorflow/tensorflow/blob/46dd63bd2c36c5b5fbdef9e2df652a5a31b55dc3/tensorflow/core/data/service/worker_impl.cc#L651-L659. PiperOrigin-RevId: 516684775
| Commit: | 7d13006 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
follow-up to cl/515714382. removes legacy trace events from the TSL Trace proto. PiperOrigin-RevId: 516321656
| Commit: | 0ec37dc | |
|---|---|---|
| Author: | Anlun Xu | |
| Committer: | Copybara-Service | |
[xla:gpu] AOT autotuning on cache miss triggers runtime autotuning PiperOrigin-RevId: 513087574
| Commit: | a290ed1 | |
|---|---|---|
| Author: | Anlun Xu | |
| Committer: | Copybara-Service | |
[xla:gpu] Add support for runtime convolution autotuning PiperOrigin-RevId: 513000771
| Commit: | f616431 | |
|---|---|---|
| Author: | Tomás Longeri | |
| Committer: | Copybara-Service | |
Replace `tensorflow::Status::SetStackTrace` with `SetStackTrace(status, trace)`, to be compatible with the `absl::Status` API. PiperOrigin-RevId: 509604863
| Commit: | 1493718 | |
|---|---|---|
| Author: | Sergey Kozub | |
| Committer: | Copybara-Service | |
Add `reordered_int8_nchw_vect` flag to convolution backend proto. This is necessary to disambiguate layouts that could not be otherwise detected by XlaConvShapesToStreamExecutorLayouts, in this case int8x32 reordered filter and bias. PiperOrigin-RevId: 508586274
| Commit: | 8f8eb4d | |
|---|---|---|
| Author: | Ilia Sergachev | |
| Committer: | Copybara-Service | |
[XLA:GPU] Add triton-based matmul emitter. PiperOrigin-RevId: 507498160
| Commit: | d8508ac | |
|---|---|---|
| Author: | Sergey Kozub | |
| Committer: | Copybara-Service | |
Add support for emitting the int8x32 cuDNN convolution reordering custom calls PiperOrigin-RevId: 506846205
| Commit: | 88acc6a | |
|---|---|---|
| Author: | Sergey Kozub | |
| Committer: | Copybara-Service | |
Add custom calls for convolution inputs reordering (cuDNN specific int8x32 layout) PiperOrigin-RevId: 506597581
| Commit: | 999e79b | |
|---|---|---|
| Author: | Ilia Sergachev | |
| Committer: | Copybara-Service | |
[XLA:GPU] Add API to query full CUDA shared memory size including the dynamic one; use it for Triton-based GEMM. PiperOrigin-RevId: 506576577
| Commit: | 4c2ec84 | |
|---|---|---|
| Author: | Ilia Sergachev | |
| Committer: | Copybara-Service | |
[XLA:GPU] Add auto-tuning of triton-based matmul. PiperOrigin-RevId: 504858571
| Commit: | 2b83de6 | |
|---|---|---|
| Author: | Ilia Sergachev | |
| Committer: | Copybara-Service | |
[XLA:GPU] Disable padding for triton-based matmuls. PiperOrigin-RevId: 504817238
| Commit: | 65c0450 | |
|---|---|---|
| Author: | Parker Schuh | |
| Committer: | Copybara-Service | |
Move xplane_to_trace_events and trace_events_to_json to tsl. PiperOrigin-RevId: 503320621
| Commit: | aecc416 | |
|---|---|---|
| Author: | Philipp Hack | |
| Committer: | Copybara-Service | |
PR #58720: FP8 GEMMs in XLA Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/58720 Enables scaled GEMMs based on `F8E4M3FN` and `F8E5M2` [FP8 data types](https://arxiv.org/abs/2209.05433). The pattern described by steps 1 through 6 in [RFC #22](https://github.com/openxla/xla/discussions/22) is rewritten into a Custom Call of the form (A, B, a_scale, b_scale, d_scale) -> (D, d_amax), where A, B and D are FP8 matrices and a_scale, b_scale and d_scale are their respective scaling factors. The scalar d_amax gives the maximum of the absolute values in D before rescaling and casting to FP8 and can be used in the calculation of new scaling factors. Copybara import of the project: -- f2eb35a9efcaaffdbb7314f99521357840bd49d8 by Philipp Hack <phack@nvidia.com>: Support for FP8 GEMMs in XLA. -- 0afd695b3840417fdb1c00987c8c5e980be0de33 by Philipp Hack <phack@nvidia.com>: Support for FP8 GEMMs in XLA. -- 5aba0882bc624215613c77d73dd23ec3b1d8b0d9 by Philipp Hack <phack@nvidia.com>: Support for FP8 GEMMs in XLA. -- 8d18d22d61b1b440421fc3dd402acdaaf27519b3 by Philipp Hack <phack@nvidia.com>: Support for FP8 GEMMs in XLA. -- 7759e0a5d041c26c632d4e433d5f544e0194ea40 by Philipp Hack <phack@nvidia.com>: Support for FP8 GEMMs in XLA. Merging this change closes #58720 PiperOrigin-RevId: 495806551
| Commit: | a70d3ac | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | Copybara-Service | |
Move profiler_options_proto to TSL PiperOrigin-RevId: 487031480
| Commit: | 9876752 | |
|---|---|---|
| Author: | Tomás Longeri | |
| Committer: | Copybara-Service | |
Roll forward of cl/478660619: Migrate core/protobuf/autotuning.proto and compiler/xla/stream_executor/dnn.proto to TSL PiperOrigin-RevId: 486269845
| Commit: | 5b3dfb6 | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | Copybara-Service | |
Move profile.proto to TSL PiperOrigin-RevId: 486230034
| Commit: | 69d51d9 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
1) Fork a minimal subset of grpc_state for grpc_coordination_client. 2) Move grpc_client_cq_tag to TSL. 3) Move proto parse/unparse methods in grpc_util from TF to TSL. 4) Add a ctor arg for RPCState to include a user-defined proto_parse_fn. 5) Use optimized tensor parse fn in worker service. PiperOrigin-RevId: 485671378
| Commit: | 9df66e3 | |
|---|---|---|
| Author: | Michael Hudgins | |
| Committer: | Copybara-Service | |
Testing failure of CI PiperOrigin-RevId: 485457160
| Commit: | 6a79f3a | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Move coordination service to TSL. PiperOrigin-RevId: 485132220
| Commit: | 09c9d6a | |
|---|---|---|
| Author: | Tomás Longeri | |
| Committer: | Copybara-Service | |
Roll forward of cl/482043512: Migrate test_log.proto to TSL Also rename test_log_proto_impl to test_log_proto and clean up loads in tsl/util/BUILD. PiperOrigin-RevId: 484965660
| Commit: | d174948 | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | Copybara-Service | |
Roll forward Move XPlane Proto to TSL. PiperOrigin-RevId: 484610094
| Commit: | d31927d | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Internal change PiperOrigin-RevId: 484280058
| Commit: | b0088b5 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Internal change PiperOrigin-RevId: 484103999
| Commit: | f94efb6 | |
|---|---|---|
| Author: | David Dunleavy | |
| Committer: | Copybara-Service | |
Duplicate python/lib/core:bfloat16_lib in XLA PiperOrigin-RevId: 484047401
| Commit: | bc58139 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Move coordination service and config protos to TSL. PiperOrigin-RevId: 484063774
| Commit: | 440c2aa | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Split RPCOptions out of config.proto and move it to TSL. PiperOrigin-RevId: 484044213
| Commit: | bd2a507 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
1.Split common methods in distributed_runtime/rpc/grpc_util to a corresponding TSL library (e.g. ToGrpcStatus). 2. Move distributed_runtime_payloads.proto to TSL. PiperOrigin-RevId: 483517874
| Commit: | 5e4705d | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | Copybara-Service | |
Migrate test_log.proto to TSL Also rename test_log_proto_impl to test_log_proto and clean up loads in tsl/util/BUILD. PiperOrigin-RevId: 482124575
| Commit: | d0e138b | |
|---|---|---|
| Author: | Tomás Longeri | |
| Committer: | Copybara-Service | |
Migrate test_log.proto to TSL Also rename test_log_proto_impl to test_log_proto and clean up loads in tsl/util/BUILD. PiperOrigin-RevId: 482043512
| Commit: | 1df87a3 | |
|---|---|---|
| Author: | Tomás Longeri | |
| Committer: | Copybara-Service | |
Migrate bfc_memory_map.proto to TSL PiperOrigin-RevId: 481833082
| Commit: | 9202127 | |
|---|---|---|
| Author: | Michael Hudgins | |
Moving original commit for tsl to preserve history
| Commit: | fa9a6f4 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | TensorFlower Gardener | |
Roll forward Move XPlane Proto to TSL. PiperOrigin-RevId: 480944931
| Commit: | 847f55e | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | TensorFlower Gardener | |
Roll forward Move XPlane Proto to TSL. PiperOrigin-RevId: 480922000
| Commit: | 6fc777b | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | TensorFlower Gardener | |
Move XPlane Proto to TSL PiperOrigin-RevId: 480418211
| Commit: | ba3e7e9 | |
|---|---|---|
| Author: | Clive Verghese | |
| Committer: | TensorFlower Gardener | |
Move XPlane Proto to TSL PiperOrigin-RevId: 480396023
| Commit: | 4ab6453 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | TensorFlower Gardener | |
Roll forward of cl/479867985: Shard core/framework/summary.proto's HistogramProto into tsl/protobuf/histogram.proto PiperOrigin-RevId: 480197757
| Commit: | 54265ed | |
|---|---|---|
| Author: | Skye Wanderman-Milne | |
| Committer: | TensorFlower Gardener | |
Migrate core/protobuf/autotuning.proto and compiler/xla/stream_executor/dnn.proto to TSL PiperOrigin-RevId: 480177205
| Commit: | 841abaf | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | TensorFlower Gardener | |
Shard core/framework/summary.proto's HistogramProto into tsl/protobuf/histogram.proto PiperOrigin-RevId: 479873997
| Commit: | 1e515b0 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | TensorFlower Gardener | |
Shard core/framework/summary.proto's HistogramProto into tsl/protobuf/histogram.proto PiperOrigin-RevId: 479867985
| Commit: | 14bb2a1 | |
|---|---|---|
| Author: | A. Unique TensorFlower | |
| Committer: | TensorFlower Gardener | |
Migrate core/protobuf/autotuning.proto and compiler/xla/stream_executor/dnn.proto to TSL PiperOrigin-RevId: 478660619