Develop upstream sync 230313 #2020

rahulbatra85 · 2023-03-15T05:42:25Z

Weekly sync

We used to remove them as part of layout normalization. Now that this is not done anymore, we should do that as part of this shape normalization. Doing it here has the benefit that we do not need to add an explicit bitcast to the HLO, as this normalization is just used for the code emitter. PiperOrigin-RevId: 514998701

There is no need to normalize broadcast operands on GPU, and it makes it harder to write matchers if they have to optionally match a bitcast. The revert is due to causing some runtime regressions PiperOrigin-RevId: 515007089

This simply makes sure that all xla binaries that are able to build are now building on presubmit in OSS. There are a few that are still failing in kokoro which I will follow up with in a follow up CL. In the future, we would also like a simple test to execute these binaries, but these targets at least assure they build in open source. PiperOrigin-RevId: 515012429

Links the forwarded attributes to an issue tracking the updates of dependent code and the evenual removal of these aliases. PiperOrigin-RevId: 515041355

Updates LLVM usage to match [e510d0bda087](llvm/llvm-project@e510d0bda087) PiperOrigin-RevId: 515051411

http://github.com/tensorflow/runtime/commit/9084a21c1fb3a8eb411a7ed010ec1e9fcaa01b01. PiperOrigin-RevId: 515054418

…rs}::Code. PiperOrigin-RevId: 515056218

…lass recently PiperOrigin-RevId: 515059613

…e conditional branches. Add cost model to trace and ensure memory pressure is not overly increased by the optimization. PiperOrigin-RevId: 515064535

PiperOrigin-RevId: 515069616

PiperOrigin-RevId: 515069653

The `experimental/kernel_tests:io_test` is about 20% flaky, and the non-experimental one is about 10% flaky. These are frequently failing in continuous builds, and sometimes blocking submissions. PiperOrigin-RevId: 515070867

They seem to have always been flaky with sanitizers, and a recent change to the llvm symbolizer makes them fail more reliably. PiperOrigin-RevId: 515074821

http://github.com/tensorflow/runtime/commit/81d91c02ded984c07857981dec74f3e60e1835f3. PiperOrigin-RevId: 515074959

…fault protocols. PiperOrigin-RevId: 515077543

PiperOrigin-RevId: 515082908

PiperOrigin-RevId: 515088268

…ops. PiperOrigin-RevId: 515089205

… Part 1: Remove obsolete operations PiperOrigin-RevId: 515089564

PiperOrigin-RevId: 515090988

PiperOrigin-RevId: 515095875

PiperOrigin-RevId: 515097373

…r original form before cl/514870586 to preserve compatibility. PiperOrigin-RevId: 515097407

PiperOrigin-RevId: 516060673

PiperOrigin-RevId: 516114140

…omposer. The preferred layout should be consistent with what we consider to be a default dot according to dot decomposer. PiperOrigin-RevId: 516125977

PiperOrigin-RevId: 516128207

PiperOrigin-RevId: 516145382

PiperOrigin-RevId: 516145402

PiperOrigin-RevId: 516146308

PiperOrigin-RevId: 516162621

Now we have a rewrite for scf.forall -> scf.for that happens earlier in the pipeline. Since there are plans to get rid of scf.parallel, this pass is not needed at all. PiperOrigin-RevId: 516199378

…ntal case. PiperOrigin-RevId: 516205387

This cl uses a preexisting grappler pass that can rewrite the batching op. It expands the rewrite support to allow rewriting batching options. Users can provide the grappler proto via the servo_tf_graph_rewrite_options flag. PiperOrigin-RevId: 516210724

Use "transformer" ops instead. PiperOrigin-RevId: 516213513

…into develop-upstream-sync-230313

rahulbatra85 · 2023-03-15T15:23:51Z

retest Ubuntu-sanity please!

i-chaochen · 2023-03-15T15:58:19Z

@rahulbatra85 Thanks for this weekly-sync!

It will be great if you could add some info on the description (i.e., fixes and disable with related tickets/PRs) like we used to do.

exemplar weekly-sync:

#1992
#1987
#2012

rahulbatra85 · 2023-03-15T15:59:53Z

@rahulbatra85 Thanks for this weekly-sync!

It will be great if you could add some info on the description (i.e., fixes and disable with related tickets/PRs) like we used to do.

exemplar weekly-sync:

#1992 #1987 #2012

Yes, I will add all the info once I am done with weekly sync. CI_Sanity and cpu run is still failing.

…ve the linkage issue when building the TF Lite kernel tests." This reverts commit 03b51b5.

rahulbatra85 · 2023-03-16T01:38:55Z

Weekly sync issues
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/3952
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/3951
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/3950

rahulbatra85 · 2023-03-16T01:41:52Z

Had to revert this f9df950. Otherwise, Ubuntu Sanity CI fails.
Tensorflow updated a commit number for dependency(google_benchmark), but it has bug around libpfm library and fails to build https://github.com/google/benchmark
A commit probably fixes the issue google/benchmark@68aa190, but it not in the v1.71 release that tensorflow moved to

akuegel and others added 30 commits March 8, 2023 04:21

Revert: Do not normalize broadcast operands on GPU.

f141bec

There is no need to normalize broadcast operands on GPU, and it makes it harder to write matchers if they have to optionally match a bitcast. The revert is due to causing some runtime regressions PiperOrigin-RevId: 515007089

Add TODO tracking for aliased attributes.

eae8207

Links the forwarded attributes to an issue tracking the updates of dependent code and the evenual removal of these aliases. PiperOrigin-RevId: 515041355

Integrate LLVM at llvm/llvm-project@e510d0bda087

67ba341

Updates LLVM usage to match [e510d0bda087](llvm/llvm-project@e510d0bda087) PiperOrigin-RevId: 515051411

Merge branch 'master' into master

11665d1

Update .github/workflows/stale-issues.yml

8ded331

Update .github/workflows/stale-issues.yml

a5d03ec

Update TFRT dependency to use revision

2c6e67c

http://github.com/tensorflow/runtime/commit/9084a21c1fb3a8eb411a7ed010ec1e9fcaa01b01. PiperOrigin-RevId: 515054418

Disable TF32 evaluation for SelfAdjointEigTest cases.

d2f4bee

Replace usage of the tsl::Status constructor with a tsl::{error, erro…

e5a23f4

…rs}::Code. PiperOrigin-RevId: 515056218

Update .github/workflows/stale-issues.yml

7aab627

Update .github/workflows/stale-issues.yml

8ef71f8

Implement FullType method in FakeTensorHandle that is added in base c…

74a9427

…lass recently PiperOrigin-RevId: 515059613

[XLA/conditional code motion] Allow reduce operands to be moved insid…

5385f06

…e conditional branches. Add cost model to trace and ensure memory pressure is not overly increased by the optimization. PiperOrigin-RevId: 515064535

Add creation moved to common elementwise creation.

24f7ee6

PiperOrigin-RevId: 515069616

Fix saved_model_cli dependencies.

9f41fc4

PiperOrigin-RevId: 515069653

Temporarily disable flaky tests.

987bea2

The `experimental/kernel_tests:io_test` is about 20% flaky, and the non-experimental one is about 10% flaky. These are frequently failing in continuous builds, and sometimes blocking submissions. PiperOrigin-RevId: 515070867

Disable sanitizers for memory_checker_test.

83ab80c

They seem to have always been flaky with sanitizers, and a recent change to the llvm symbolizer makes them fail more reliably. PiperOrigin-RevId: 515074821

Update TFRT dependency to use revision

811dd0c

http://github.com/tensorflow/runtime/commit/81d91c02ded984c07857981dec74f3e60e1835f3. PiperOrigin-RevId: 515074959

Support falling back to gRPC at GetElement time for experimental de…

e2c2dc4

…fault protocols. PiperOrigin-RevId: 515077543

Merge pull request tensorflow#59594 from tensorflow:pjpratik-patch-2

b17ed11

PiperOrigin-RevId: 515082908

Update GetBlasComputationType

b48aa62

[xla:gpu] Serialize runtime convolution autotuning

763e78d

PiperOrigin-RevId: 515088268

Don't drop frontend attributes during MHLO<->HLO translation in some …

d4eb69d

…ops. PiperOrigin-RevId: 515089205

Rollforward of PR tensorflow#58640: [INTEL oneDNN] OneDNN Refactoring…

155b34b

… Part 1: Remove obsolete operations PiperOrigin-RevId: 515089564

Encapsulate FuncGraph _capture_helper logics in capture_container

adc7074

PiperOrigin-RevId: 515090988

Propagate cost threshhold to parallelization pass.

1e34bfb

PiperOrigin-RevId: 515095875

[MLIR] Add deallocation dialect to the bisect tool

4d555d5

PiperOrigin-RevId: 515097373

Revert the signature of Logical{And,Or,Not} ops in TensorFlow to thei…

08c537a

…r original form before cl/514870586 to preserve compatibility. PiperOrigin-RevId: 515097407

ashishenoyp and others added 19 commits March 12, 2023 15:19

Fix memory leak in TensorInterfaceFromTensor().

1ab3b73

PiperOrigin-RevId: 516060673

Add support for VHLO in XlaCallModuleOp and JAX2TF

fe110a1

PiperOrigin-RevId: 516114140

Choose the same preferred layout for the first dot operand as dot dec…

d0855e8

…omposer. The preferred layout should be consistent with what we consider to be a default dot according to dot decomposer. PiperOrigin-RevId: 516125977

[XLA] Add CollectivePermute op support to pattern matcher.

91824fe

PiperOrigin-RevId: 516128207

Update GraphDef version to 1434.

a83a3a7

PiperOrigin-RevId: 516145382

compat: Update forward compatibility horizon to 2023-03-13

b2737c9

PiperOrigin-RevId: 516145402

Adjust failing tests, now scf.forall is rewritten to scf.for.

fbb5020

PiperOrigin-RevId: 516146308

split-alloc-tensors: Don't crash for empty regions.

3693f68

PiperOrigin-RevId: 516162621

[GmlSt] Remove scf.forall -> scf.parallel pass.

3ad732b

Now we have a rewrite for scf.forall -> scf.for that happens earlier in the pipeline. Since there are plans to get rid of scf.parallel, this pass is not needed at all. PiperOrigin-RevId: 516199378

Check data transfer client<-server compatibility only in the experime…

7f577b6

…ntal case. PiperOrigin-RevId: 516205387

[GmlSt] Remove tiling pass that uses labels to target the op to tile.

0da98d2

Use "transformer" ops instead. PiperOrigin-RevId: 516213513

Merge(unresolved conflicts) remote-tracking branch 'upstream/master' …

cc8ac6e

…into develop-upstream-sync-230313

Resolved Conflicts for 230313 sync

c5c1bb6

[ROCm]: Fix amdgcn hlo tests

f58dd21

[ROCm]: Disable BF16 test in XLA matrix ops simple test

ff4849a

[ROCm]: Disable Xla tools multihost hlo runner test

a8176dc

[ROCm]: Disable new kernel reuse test on ROCm

839480b

[ROCm]: Don't run multi client test during single gpu run

4aedd33

Rahul Batra added 2 commits March 15, 2023 20:09

Revert "Update the com_google_benchmark dependency to v1.7.1 to resol…

f9df950

…ve the linkage issue when building the TF Lite kernel tests." This reverts commit 03b51b5.

[ROCm]: Don't run xla/gpu hlo_to_llvm_ir build test on CPU

6f94cab

rahulbatra85 requested a review from i-chaochen March 16, 2023 01:39

rahulbatra85 requested a review from jayfurmanek March 16, 2023 02:10

i-chaochen approved these changes Mar 16, 2023

View reviewed changes

rahulbatra85 merged commit 3ad95d0 into develop-upstream Mar 16, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Develop upstream sync 230313 #2020

Develop upstream sync 230313 #2020

Uh oh!

rahulbatra85 commented Mar 15, 2023

Uh oh!

rahulbatra85 commented Mar 15, 2023

Uh oh!

i-chaochen commented Mar 15, 2023 •

edited

Loading

Uh oh!

rahulbatra85 commented Mar 15, 2023

Uh oh!

rahulbatra85 commented Mar 16, 2023

Uh oh!

rahulbatra85 commented Mar 16, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

72 participants

Develop upstream sync 230313 #2020

Develop upstream sync 230313 #2020

Uh oh!

Conversation

rahulbatra85 commented Mar 15, 2023

Uh oh!

rahulbatra85 commented Mar 15, 2023

Uh oh!

i-chaochen commented Mar 15, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rahulbatra85 commented Mar 15, 2023

Uh oh!

rahulbatra85 commented Mar 16, 2023

Uh oh!

rahulbatra85 commented Mar 16, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

72 participants

i-chaochen commented Mar 15, 2023 •

edited

Loading