Develop upstream sync 240521 #2548
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
PiperOrigin-RevId: 634161496
…KindInfo()` Checking `GetEnableMemories()` inside `PyDeviceList::PopulateMemoryKindInfo()` causes the memory kind to be built/cached based on the value of `enable_memories` at the time the method was called for the first time. This is problematic when users later change the value of `enable_memories` but reuse the same device list. This CL fixes the problem by making `PyDeviceList::MemoryKinds()` and `PyDeviceList::DefaultMemoryKind()` always check for `GetEnableMemories()` outside the cached method to make it behave correctly even if `enable_memories` changes. PiperOrigin-RevId: 634191502
PiperOrigin-RevId: 634193569
PiperOrigin-RevId: 634195071
`xla::ifrt::RemapPlan`'s `mappings` is frequently reused across different `xla::ifrt::RemapPlan`. It can be costly to keep a copy of `mappings` for each plan if the mappings is not trivially small. This change wraps `mappings` with `std::shared_ptr` so that it can be shared more cheaply across multiple plans. PiperOrigin-RevId: 634214931
PiperOrigin-RevId: 634216731
PiperOrigin-RevId: 634224863
… tests. Imported from GitHub PR openxla/xla#12538 Copybara import of the project: -- 78d74154a1693426014eef2940e759d1e36902ed by Ilia Sergachev <[email protected]>: [GPU] Fix OSS compilation of previously disabled tests. Merging this change closes tensorflow#12538 PiperOrigin-RevId: 634248275
PiperOrigin-RevId: 634257201
The previous logic for filtering by-module wasn't correct, as instructions could be modified after autotuning, resuling in not all relevant information serialized. This could result in excessive data serialized, but autotuning results are very small compared to the module size and the compiled artifact. PiperOrigin-RevId: 634259538
…mRewriter pass. PiperOrigin-RevId: 634263367
…eVars PiperOrigin-RevId: 634264428
PiperOrigin-RevId: 634268437
PiperOrigin-RevId: 634268459
XLA should not include any platform specific headers in XLA header files. This change moves us one step closer to this goal by getting rid of type aliases in `gpu_types.h`. PiperOrigin-RevId: 634278649
PiperOrigin-RevId: 634279450
PiperOrigin-RevId: 634285627
Imported from GitHub PR openxla/xla#12501 Copybara import of the project: -- 97b50d47877dd3dc46535c6fa34f0e449ee57bfe by Ilia Sergachev <[email protected]>: [NFC] Fix mistypes in scatter expander. Merging this change closes tensorflow#12501 PiperOrigin-RevId: 634285954
PiperOrigin-RevId: 634288224
PiperOrigin-RevId: 634291787
…fusion_analysis. This is a pure HLO pass that does not require a GPU. PiperOrigin-RevId: 634303515
This is used as an input to mlir_replay. Note that GPU-specific ops are not implemented in the interpreter yet (at HEAD). PiperOrigin-RevId: 634304231
…keep track of the cache path. PiperOrigin-RevId: 634308967
…NN_3.4.1 PiperOrigin-RevId: 634320860
For this to work, we need a folder, so add that too. We kind of rely on the folder to work most of the time (for caching of instructions). There may be some dragons here, I'll probably rewrite that stuff at some point. This works around the remaining known miscompiles. The issue must be somewhere in LLVM and be related to GEPs with subtractions, but I lack the skills and resources to actually track it down. PiperOrigin-RevId: 634326550
PiperOrigin-RevId: 634330969
…he module. PiperOrigin-RevId: 634349913
…m_algorithm_picker. These defines are not needed since the BUILD target is anyway guarded by if_gpu_is_configured. PiperOrigin-RevId: 634353818
|
https://clang.llvm.org/docs/ThreadSafetyAnalysis.html --config=warnings? |
|
@draganmladjenovic I think you will upstream this profiler as well later? |
i-chaochen
reviewed
May 29, 2024
There was a problem hiding this comment.
We could try to enable this cudnn_fused_conv_rewriter_test this week as openxla/xla#9666 is merged
|
retest Ubuntu-GPU-multi please |
Nope. Still fails. |
draganmladjenovic
force-pushed
the
develop-upstream-sync-240521
branch
2 times, most recently
from
4e1e7ee to
f1d0cc8
Compare
|
@draganmladjenovic let's disable this one as well so this weekly-sync should be ok. We should enable some of disable XLA ones next weekly-sync. |
draganmladjenovic
force-pushed
the
develop-upstream-sync-240521
branch
2 times, most recently
from
bae8034 to
161cea2
Compare
|
retest Ubuntu-GPU-single please |
To avoid loading both system and locald_rocm_config version of the library. WORKAROUND!
draganmladjenovic
force-pushed
the
develop-upstream-sync-240521
branch
from
161cea2 to
4e0d9b8
Compare
i-chaochen
reviewed
Jun 3, 2024
There was a problem hiding this comment.
@Ruturaj4 @rahulbatra85 please check this on JAX side.
draganmladjenovic
force-pushed
the
develop-upstream-sync-240521
branch
from
4e0d9b8 to
cc4a257
Compare
|
retest Ubuntu-GPU-single please |
1 similar comment
|
retest Ubuntu-GPU-single please |
|
retest Ubuntu-CPU please |
…t/cost_analysis.h
|
retest Ubuntu-GPU-single please |
|
retest Ubuntu-GPU-single please |
|
@i-chaochen Please review. |
i-chaochen
approved these changes
Jun 5, 2024
There was a problem hiding this comment.
@draganmladjenovic could you upstream this profile change? 4736848
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.