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.

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

Develop upstream sync 230515 #2106

Merged
merged 349 commits into from
May 19, 2023

Conversation

wenchenvincent
Copy link

No description provided.

MarkDaoust and others added 30 commits May 9, 2023 14:53
PiperOrigin-RevId: 530719240
PiperOrigin-RevId: 530729537
Updates LLVM usage to match
[e9569748def8](llvm/llvm-project@e9569748def8)

PiperOrigin-RevId: 530732122
We check whether the PjRtClient has changed or not before compilation. We delete and create a new PJRT DeviceCompiler if it has changed.

Ideally the cache should be invalidated upon shutdown, although getting a hold of the TpuDevice in TpuSystemShutdown() doesn't seem straightforward (the cache is stored in the TpuDevice's resource manager).

PiperOrigin-RevId: 530751696
…more clear that callers should pass them explicitly.

PiperOrigin-RevId: 530752496
…e_id_tensor_int32_t

PiperOrigin-RevId: 530776070
…evice information.

The previous implementation was non-deterministic, as it relied on absl
traversal order and functional control flow operations having a particular
arrangement of nodes.

PiperOrigin-RevId: 530781619
…t_id` helper.

Compression is an input to `register_dataset`, not `from_dataset_id`.

PiperOrigin-RevId: 530783906
PiperOrigin-RevId: 530795290
…p_outside_compilation can be used in the SPMD case and there are num_cores_per_replica host regions. Host regions are located on the host near the corresponding TPU core. Host regions are created by cloning the first region.

A tensor that is sent or received to the host is split sharded. This tensor needs to be converted to MANUAL sharding before send/recv with _XlaHostComputeMlir.

PiperOrigin-RevId: 530801371
This uses an ApproxTopK custom-call, which we add support for in supported by
MHLO, by including a lowering to XLA's PartialReduce custom_call via the Client
XLA ApproxTopK function.

PiperOrigin-RevId: 530805966
See openxla/xla#2845

The modified code triggered:
`warning C5101: use of preprocessor directive in function-like macro argument list is undefined behavior`

Due to macro guards inside of a function-like macro invocation.

The solution here was to merely clone the macro call into each directive
region.

PiperOrigin-RevId: 530835539
PiperOrigin-RevId: 530845249
- use-after-free
- double-free
- detect more out of bounds accesses

PiperOrigin-RevId: 530847579
PiperOrigin-RevId: 530855903
Use optimized 4bit eval when weight shape is larger than [4x32]

PiperOrigin-RevId: 530859000
wilsingosti and others added 26 commits May 15, 2023 08:32
… experiment to debug why some jobs got slower.

PiperOrigin-RevId: 532120751
PiperOrigin-RevId: 532132008
…sformers/ have strict dependencies.

PiperOrigin-RevId: 532136646
PiperOrigin-RevId: 532138742
PiperOrigin-RevId: 532141388
…ta/experimental/kernel_tests/optimization/ have strict dependencies.

PiperOrigin-RevId: 532143351
PiperOrigin-RevId: 532148739
PiperOrigin-RevId: 532153784
PiperOrigin-RevId: 532161692
When one worker has finished, the snapshot manager is in the
"WindingDown" state. It doesn't need to create new streams.

If it does, a worker is assigned two streams.

PiperOrigin-RevId: 532171233
This helps with better visualizing Triton GEMMs as HLO ops in profiling tools

PiperOrigin-RevId: 532173667
PiperOrigin-RevId: 532176002
PiperOrigin-RevId: 532178377
…rotos from AbstractStackTrace and StackTracesMap objects. Its API has options that cover various typical use cases.

PiperOrigin-RevId: 532183689
@wenchenvincent wenchenvincent merged commit 5e47f2b into develop-upstream May 19, 2023
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet