[pull] master from tensorflow:master - #8799
Merged
Merged
Conversation
Bumps [markdown](https://github.com/Python-Markdown/markdown) from 3.8 to 3.8.1. - [Release notes](https://github.com/Python-Markdown/markdown/releases) - [Changelog](https://github.com/Python-Markdown/markdown/blob/master/docs/changelog.md) - [Commits](Python-Markdown/markdown@3.8...3.8.1) --- updated-dependencies: - dependency-name: markdown dependency-version: 3.8.1 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com>
…-data-service-server-lib-crash
This API is not used by custom kernels anymore. Right now it contains only cross-device barrier related logic. PiperOrigin-RevId: 9744425
Overrides GetDebugOptionsForTest to set xla_gpu_experimental_gemm_fusion_v2 to false for HoistFusedBitcastsReshapeTest. We don't run HoistFusedBitcasts when gemm_fusion_v2=true, so this test doesn't make sense for it. Set it to false explicitely. PiperOrigin-RevId: 974445800
This reduces the number of kernels from 4 to 2 in the following example:
```
// Input HLO:
%p = f16[32] parameter(0)
%b = f16[1,9,9,32] broadcast(%p), dimensions={3}
%add1 = add(%conv1, %b)
%add2 = add(%conv2, %b)
// Before (4 separate kernels emitted):
%conv1_out = convolution(...)
%add1_out = fusion(%conv1_out, %b), kind=kLoop
%conv2_out = convolution(...)
%add2_out = fusion(%conv2_out, %b), kind=kLoop
// After (2 fused cuDNN graph kernels, 0 intermediate HBM traffic):
%fusion1 = fusion(%input1, %filter1, %p), kind=kCustom, calls={
%p0 = f16[32] parameter(2)
%b0 = f16[1,9,9,32] broadcast(%p0), dimensions={3}
ROOT %out = add(%conv1, %b0)
}
%fusion2 = fusion(%input2, %filter2, %p), kind=kCustom, calls={
%p0 = f16[32] parameter(2)
%b0 = f16[1,9,9,32] broadcast(%p0), dimensions={3}
ROOT %out = add(%conv2, %b0)
}
```
PiperOrigin-RevId: 974473390
Unless we run with `mark_fusion_duplications = true`, we should just ignore fusion computations. We are not allowed to deduplicate them. PiperOrigin-RevId: 974488637
PiperOrigin-RevId: 974490864
…& V2 in preparation of V2 launch. Remove test MultipleBatchRequireSeparateTranspose - this is a fusion heuristic that does not belong here and is different than the new transpose fusion heuristic that we chose for V2. Rewrite LinkingLibdeviceTwiceWorks - this test is meant to confirm that a module with multiple triton gemms is able to launch, but the matcher is too dependent on fusion decisions that are not relevant for the test. As mentioned in the TODO, it would be better if this was rewritten to post-fusion & only run once Triton was chosen for these 2 gemms so that's what I did. PiperOrigin-RevId: 974560034
PiperOrigin-RevId: 974570161
PiperOrigin-RevId: 974572981
…/requirements_updater/numpy1_requirements/markdown-3.8.1 PiperOrigin-RevId: 974581495
…tal_gemm_fusion flag in each case that it's meant to represent. These will get cleaned up after the flag flip, but explicitly setting it means that the tests themselves are not affected (they run both versions when they're meant to or only 1 version where they differ intentionally). PiperOrigin-RevId: 974584295
When an optimized remap plan provides `input_devices_for_output_map`, it explicitly details which input arrays and device shards contribute to each output array. Previously, `PjRtCompatibleClientRemapArrays` always inspected `plan.mappings()` regardless of whether `input_devices_for_output_map` was present. Update `PjRtCompatibleClientRemapArrays` to construct output buffer assignments directly from `plan.input_devices_for_output_map()` when available, bypassing `plan.mappings()` traversal. Enforce in `RemapPlan::Validate` that `input_devices_for_output_map`, if specified, must be fully populated (`size() == num_outputs`). PiperOrigin-RevId: 974584311
…usion_analysis_test.cc PiperOrigin-RevId: 974585794
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
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
8 participants
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.
See Commits and Changes for more details.
Created by
pull[bot] (v2.0.0-alpha.4)
Can you help keep this open source service alive? 💖 Please sponsor : )