Tfci debug part2 #11
Open
Tfci debug part2 #11
Conversation
This file contains hidden or 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: 756853517
…ngs::StrAppend`. `strings::StrCat` should eventually forward to `absl::StrCat`. Some references need to be rewritten as `absl::StrCat(absl::LegacyPrecision(...))` to avoid loss of precision. PiperOrigin-RevId: 756865936
PiperOrigin-RevId: 756867042
…ation environments PiperOrigin-RevId: 756881629
`.size()` and `operator[]` have a race condition. With this fix, the threads won't access the container itself; they just writes the elements they need to modify. PiperOrigin-RevId: 756902174
…tion given. Currently, you have to specify the return type on ArrayTypeSwitch, but it is often redundant as it can be inferred from the functor provided. PiperOrigin-RevId: 756909206
…t::Compile` before MLIR -> XlaComputation conversion PjRt GPU applies additional transformations to preserve input/output layout, which is only implemented in `StreamExecutorGpuClient::CompileAndLoad` and not in `StreamExecutorGpuCompiler::Compile`. Doing the MLIR -> XlaComputation conversion without this logic causes custom layouts to be dropped. PiperOrigin-RevId: 756914304
CpuClient::CreateUninitializedBuffer. PiperOrigin-RevId: 756916747
PiperOrigin-RevId: 756921029
The version of `upb` used in tensorflow and XLA is incompatible with Clang. In particular, it generates a warning that the code uses a non-standard C++ feature. Since this version of `upb` has `-Werror` in its build opts, the warning breaks the build. We want to be able to compile PyTorch/XLA with clang, and PyTorch/XLA depends on `upb`. Therefore we need to make `upb` buildable with Clang. In this change, we remove `-Werror` from `upb`'s build opts to prevent the warnings generated by Clang to break the build. In general, we should never use `-Werror` on code that we don't directly control, as our ability for fixing the warnings in such code is limited. PiperOrigin-RevId: 756927870
We should never crash when printing an XLA construct, even when it's invalid. PiperOrigin-RevId: 756968212
PiperOrigin-RevId: 756980155
…ow the initial fusion worklist is formed for a current computation. PiperOrigin-RevId: 756997596
…fs using `tsl::SerializeToStringDeterministic` PiperOrigin-RevId: 757021740
PiperOrigin-RevId: 757035793
PiperOrigin-RevId: 757044032
PiperOrigin-RevId: 757044520
PiperOrigin-RevId: 757054230
default memory type. The configuration option `legacy_memory_space_behavior`, which currently defaults to true, controls whether the old or the new behavior is followed. PiperOrigin-RevId: 757057567
PiperOrigin-RevId: 757084827
PiperOrigin-RevId: 757085928
…attening. PiperOrigin-RevId: 757090767
PiperOrigin-RevId: 757091461
…mic shape The previous CL that changed this logic to use the on-device shape from the device buffer generates literals with invalid sizes when the PjRt buffer has a dynamic shape. PiperOrigin-RevId: 757091876
PiperOrigin-RevId: 757110017
PiperOrigin-RevId: 757152089
PiperOrigin-RevId: 757181155
PiperOrigin-RevId: 757198443
PiperOrigin-RevId: 757209184
With Triton multi-output fusions, we can have tuple results for fusions. Adjust the buffer comparison logic accordingly. PiperOrigin-RevId: 759002766
PiperOrigin-RevId: 759010423
PiperOrigin-RevId: 759045624
PiperOrigin-RevId: 759046188
PiperOrigin-RevId: 759046243
Also fix the BUILD file, so we do not skip testing this on H100. PiperOrigin-RevId: 759050327
propagate broadcast multiplier upwards through all the ops up to the parameter. The broadcast adds a new dim or with a bitcast expand an old dim. When the expansion happens we set the broadcast multiplier to the source instruction. But as of now if before the broadcast we have more than one instruction we reset the broadcast multiplier back to one. Lets not do that. PiperOrigin-RevId: 759051178
PiperOrigin-RevId: 759054070
PiperOrigin-RevId: 759068702
We weren't handling them correctly meaning you couldn't use a `shard_map`/`ManualComputationOp` which has callbacks inside. PiperOrigin-RevId: 759072597
The autotuner compile util does not run any HLO passes. So disabling the triton softmax pass is a no-op. We get rid of the triton fusion by taking the fusion computation and running just some dedicated passes (like PriorityFusion). PiperOrigin-RevId: 759075358
PiperOrigin-RevId: 759075627
PiperOrigin-RevId: 759082316
We fixed the underlying issue with the subchannel dequantize ops sequence like below: param->transpose->broadcast->bitcast->multiply->dot Now we could remove the flag-flip from the tests PiperOrigin-RevId: 759095250
PiperOrigin-RevId: 759095404
…ne_parallelism_opt_level PiperOrigin-RevId: 759104549
PiperOrigin-RevId: 759106677
…y on zero termination There's nothing guaranteeing that these references are terminated. PiperOrigin-RevId: 759111351
PiperOrigin-RevId: 759112955
PiperOrigin-RevId: 759118994
PiperOrigin-RevId: 759127256
PiperOrigin-RevId: 759133213
…_device_test.cc PiperOrigin-RevId: 759133817
This should make the swizzle mode more readable and does not mislead the reader by having it set to "0" before it has even considered which one it should be doing. swizzle_mode is specifically left out when it is unset. PiperOrigin-RevId: 759134652
PiperOrigin-RevId: 759135428
PiperOrigin-RevId: 759137056
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.