Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Develop upstream sync 231030 #2287

Merged
merged 247 commits into from
Nov 7, 2023
Merged

Develop upstream sync 231030 #2287

merged 247 commits into from
Nov 7, 2023

Conversation

rahulbatra85
Copy link

No description provided.

bartchr808 and others added 30 commits October 23, 2023 04:44
@bartchr808 @tensorflower-gardener
Print optional attr-dict on stablehlo.reduce.
dc28c32 
In the compact version of printing, any custom attributes a user of StableHLO sets will not appear. This makes sure they appear if they exist.

PiperOrigin-RevId: 575778539
@apivovarov @tensorflower-gardener
PR tensorflow#6475: Fix flaky test bitcast_dtypes_expander_test
a39d7df 
Imported from GitHub PR openxla/xla#6475

`bitcast_dtypes_expander_test` failed when executed in Docker container
```
//xla/service:bitcast_dtypes_expander_test                               FAILED in 3 out of 3 in 0.7s
```

Reason - swapped variable names in `S64toS32` test
```
%shift-right-logical.11 -> %shift-right-logical.12
%constant.12 -> %constant.11
```
To resolve the issue we can use pattern based var names instead.
```
%[[VAL_10:.*]]
%[[VAL_11:.*]]
```

Copybara import of the project:

--
b99ed813b39c6b53f1aa0981dbaae060d8cb5aa8 by Alexander Pivovarov <pivovaa@amazon.com>:

Fix flaky test bitcast_dtypes_expander_test

Merging this change closes tensorflow#6475

PiperOrigin-RevId: 575793343
@d0k @tensorflower-gardener
Integrate LLVM at llvm/llvm-project@e558be51bab0
6f30fc1 
Updates LLVM usage to match
[e558be51bab0](llvm/llvm-project@e558be51bab0)

PiperOrigin-RevId: 575794929
@MichaelHudgins @tensorflower-gardener
Add libcublas-12-2 to prevent the libnvinfer8 dependency from pulling…
a02bae3 
… cuda 12.3 dependencies.

Otherwise, when pulled, the 12.3 will change the alias of /usr/local/cuda to 12.3, causing failures.

PiperOrigin-RevId: 575799576
@gnecula @tensorflower-gardener
[XlaCallModule] Add more logging.
c251034 
Log the op attributes on `--vmodule=xla_call_module_op=3`.

PiperOrigin-RevId: 575812286
@matthiaskramm @tensorflower-gardener
Introduce tf.GeneratorDatasetRegion op.
3441b35 
PiperOrigin-RevId: 575813013
@tensorflower-gardener
Update TFRT dependency to use revision
1b57ad5 
http://github.com/tensorflow/runtime/commit/4e2f93c2c98c3ee32fd62ef7dbd5cc75c80011b8.

PiperOrigin-RevId: 575813261
@tensorflower-gardener
[XLA:GPU] Avoid scheduling already scheduled modules.
ba09559 
PiperOrigin-RevId: 575813414
@tdanyluk @tensorflower-gardener
[XLA:GPU] Handle NaNs correctly in minimum and maximum
68e88d2 
Fixed that IrEmitterTriton "swallowed" NaNs in minimum and maximum.

Fixed that ElementalIrEmitter "swallowed" NaNs in the case of minimum(x, NaN) on GPU. That was likely caused by an llvm error.

PiperOrigin-RevId: 575817645
@fergushenderson @tensorflower-gardener
Add forwarding shim for the C++ XNNPack plugin.
c3a5443 
PiperOrigin-RevId: 575825788
@rocketas @tensorflower-gardener
Add boilerplate code to add verify clustering passes.
729644e 
PiperOrigin-RevId: 575845487
@tensorflower-gardener
Update TFRT dependency to use revision
319d697 
http://github.com/tensorflow/runtime/commit/18a68a86884cc3b033586e725cb65557420fe3ea.

PiperOrigin-RevId: 575847883
@tensorflower-gardener
Touch up the requirements' updater README.
4f36d36 
PiperOrigin-RevId: 575850088
@anlunx @tensorflower-gardener
[xla:gpu] Allow users to toggle command types that are enabled in com…
ca42872 
…mand buffers

Before this CL the libraries supported by GPU graphs are determined a integer level. This CL allows the user to have fine-grained control on whether a library should be enabled in graphs.

Example usage: XLA_FLAGS=--xla_gpu_command_buffer_command_types=FUSION,CUBLAS

PiperOrigin-RevId: 575856977
@tensorflower-gardener
Refactor Requantize in ConvertMHLOQuantToInt pass
1ce1300 
Consolidate implementations of requantize and use the same one as TF quantizer.

PiperOrigin-RevId: 575864178
@tensorflower-gardener
Removed the unused device argument.
d734e0d 
PiperOrigin-RevId: 575866461
@tensorflower-gardener
[XLA:Runtime] Moved the FFT thunk to a new folder and removed unused …
affb9df 
…dependencies, as part of a thunk clean up, and updated the necessary directories pointing to this thunk. tensorflow#5758

PiperOrigin-RevId: 575867727
@shan3290 @tensorflower-gardener
Make costs of each batch accessible to RPC handler.
360b6e3 
PiperOrigin-RevId: 575879128
@tensorflower-gardener
Add a 3.12 Docker container.
a9e19d0 
PiperOrigin-RevId: 575881215
@tensorflower-gardener
Logs a warning if no default sharding strategy can be found for a giv…
f2a52bd 
…en instruction, and reports the total # of such nodes.

PiperOrigin-RevId: 575887322
@jimlinntu @tensorflower-gardener
#tf-data destroys element before the mutex_lock l(*mu) is destroyed.
d5d0dee 
This will guarantee that the iterators created inside of parallel interleave iterator get destroyed before itself.
Otherwise, `CancelThreads(true)` might finish before `element`'s destructor is called because of `outstanding_threads_finished_cond_var_.wait(l);`

PiperOrigin-RevId: 575888988
@chr1sj0nes @tensorflower-gardener
[stream_executor] Use StreamIsCapturing function.
4fd71da 
PiperOrigin-RevId: 575890152
@mrry @tensorflower-gardener
Remove dependency from "GetMinibatchSplitsWithPhysicalReplica" to a f…
9e31959 
…lag.

The flag could have a different value in different processes, which would invalidate the inferred shape. The dependency also inflates the size of the op library, by adding a dependency on a kernel library (and transitively on MLIR), and pulls in unrelated ops which complicates wrapper generation.

PiperOrigin-RevId: 575900618
@tensorflower-gardener
Fixes typo in function args comment
1c22230 
PiperOrigin-RevId: 575904949
@changm @tensorflower-gardener
Move error logging into lower cluster to runtime ops
10ac33a 
PiperOrigin-RevId: 575910019
@ddunl @tensorflower-gardener
Refactor xla.bzl to not repeat deps list in xla_cc_{binary,test}
0ed832a 
In preparation to create `xla_protos_all` to simplify further

PiperOrigin-RevId: 575911998
@mrry @tensorflower-gardener
Remove dependencies on other op libraries from //third_party/tensorfl…
57b76a1 
…ow/core/tpu/ops:sparse_core_ops.

These dependencies are unnecessary, and they can introduce duplicate op wrappers when we generate Python bindings.

PiperOrigin-RevId: 575921758
@tensorflower-gardener
Add visibility for //learning/metadata/artifactoid/cc
2328ddf 
PiperOrigin-RevId: 575932165
@tensorflower-gardener
Sprawling .pyi updates related to pybind11 PRs tensorflow#4876.
be859e0 
PiperOrigin-RevId: 575941719
@angerson @tensorflower-gardener
Add nightly libtpu download to newer scripts
cb1f995 
Trying to get these in sync with the current jobs, which are failing,
probably because of the wrong .so being used.

PiperOrigin-RevId: 575978433
tdanyluk and others added 14 commits October 30, 2023 11:43
@tdanyluk @tensorflower-gardener
[XLA:GPU] Make the order of Executable candidates deterministic for e…
8f8efc6 
…ach fusion

This is needed for xla_gpu_deterministic_ops to function reliably - as tested by determinism_test.cc.

PiperOrigin-RevId: 577906200
@tensorflower-gardener
Re-enable layering_check for package.
5f91321 
PiperOrigin-RevId: 577906242
@tensorflower-gardener
Fix highway c++ library includes from highway/hwy/base.h -> hwy/base.h
e581ec3 
PiperOrigin-RevId: 577907744
@ddunl @tensorflower-gardener
Fix formatting of copybara:uncomment directives to help fix rocm build
c0ea018 
PiperOrigin-RevId: 577915672
@tensorflower-gardener
[TF:NPD] Support PjRtTensorBuffer in TF_CreatePjRtBuffer (which calls…
73628e5 
… SetPjRtCBufferToTensor).

PiperOrigin-RevId: 577915709
@cliveverghese @tensorflower-gardener
Fix Trace Viewer pid numbering
b059752 
PiperOrigin-RevId: 577915784
@tensorflower-gardener
Adds CreateTensorProtoSpan(), an alternative to
e902413 
CreateTensorProto() that allows data to be loaded from
non-std::vector sources, which is often more efficient.

vector<bool> makes the implementation of internal::CreateTensorProto
a bit messier than it would otherwise be due to it not being
representable as an absl::Span.

PiperOrigin-RevId: 577926897
@ezhulenev @tensorflower-gardener
[xla:gpu] Add user-managed workspace to legacy cuBLAS calls
6c2a8fc 
In StreamExecutor we rely on a thread local hack to pass workspace to cuBLAS, long term we need to make workspace explicit in all BLAS APIs.

Without user-managed workspace cuBLAS adds memory allocation nodes when captured into CUDA graphs, and that triggers graph update bugs in CUDA.

PiperOrigin-RevId: 577928686
@fergushenderson @tensorflower-gardener
Add forwarding shim for the C++ XNNPack plugin.
56a269d 
PiperOrigin-RevId: 577935297
@majiddadashi @tensorflower-gardener
[tflite] Add stablehlo.multiply reference kernel to tflite.
5c5f0fa 
PiperOrigin-RevId: 577941102
@tdanyluk @tensorflower-gardener
[XLA:GPU] Modify autotuning configs based on the input/output size of…
8f1473f 
… the dot

1. Reduce block sizes if they are bigger than NextPowerOf2(tensor size).
2. Reduce split_k to NextPowerOf2(k)/block_k.
3. Remove duplicates created by the previous steps.

I intend this cl not to make any performance degradation.

PiperOrigin-RevId: 577941843
@tensorflower-gardener
Change the OSS visibility of batch_op_rewriter_proto
47dcda3 
PiperOrigin-RevId: 577942475
@jurahul @tensorflower-gardener
[XLA/GPU] Add debug option to set the slop factor for memory limit
44bb031 
PiperOrigin-RevId: 577946530
Merge(Conflicts) remote-tracking branch 'upstream/master' into develo…
abf7b02 
…p-upstream-sync-231030
@rahulbatra85 rahulbatra85 force-pushed the develop-upstream-sync-231030 branch from 669ff45 to 34f29b8 Compare November 2, 2023 18:03
@rahulbatra85
Copy link
Author

Unit test failures
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/6261
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/6260
https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/6264

i-chaochen
i-chaochen reviewed Nov 3, 2023
View reviewed changes
tensorflow/compiler/tests/xla_call_module_test.py Outdated
@@ -496,7 +497,7 @@ def f(x):
}
}
"""

'''
Copy link

@i-chaochen i-chaochen Nov 3, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you use test.is_built_with_rocm() instead of just commented out in the code? Also, in your tracking ticket https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/6264 could you also add the new added commit from upstream for tracking? Thanks

Rahul Batra added 3 commits November 3, 2023 16:03
Resolved conflicts for develop-upstream-sync-231030
d165aed
[ROCm] fixing static linkage problems
8dd0424
[ROCm]: Disable unit tests(231030 sync)
0c62bc3 
	//tensorflow/compiler/tests:xla_call_module_no_platform_check_test_gpu
	//tensorflow/compiler/tests:xla_call_module_no_shape_assertions_check_test_gpu
	//tensorflow/compiler/tests:xla_call_module_test_gpu
	//tensorflow/python/eager:context_test_gpu
	//tensorflow/python/ops/memory_tests:custom_gradient_memory_test_gpu
@rahulbatra85 rahulbatra85 force-pushed the develop-upstream-sync-231030 branch from 34f29b8 to 0c62bc3 Compare November 3, 2023 16:06
@rahulbatra85
Copy link
Author

retest Ubuntu-CPU please!

1 similar comment
@rahulbatra85
Copy link
Author

rahulbatra85 commented Nov 3, 2023 •
edited
Loading

retest Ubuntu-CPU please!

@i-chaochen
Copy link

retest Ubuntu-CPU please

@rahulbatra85 rahulbatra85 merged commit 5351bd5 into develop-upstream Nov 7, 2023
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jayfurmanek jayfurmanek jayfurmanek approved these changes

@i-chaochen i-chaochen i-chaochen approved these changes

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.