[pull] master from tensorflow:master#587
Merged
pull[bot] merged 22 commits into3M-com:masterfrom Sep 24, 2025
Merged
Conversation
PiperOrigin-RevId: 810523749
Recall that when a task in a multi-controller JAX (mcjax) job starts, it is
given a randomly generated incarnation id. If the task fails and restarts, it
gets a new id. These incarnation ids help us distinguish two different
instances of the same task.
When a group of tasks wants to perform a collective (e.g., AllReduce), every
participate needs to agree on the set of participants. If task 0 thinks the
collective is with tasks {0, 1, 2} while task 1 thinks the collective is with
tasks {0, 1, 3}, the program will not execute smoothly.
The participants also need to agree on everyone's incarnations. For example, if
task 0 thinks task 2 has incarnation 42 while task 1 thinks task 2 has
incarnation 100, they will disagree on whether task 2 has failed or not.
Tasks get the set of participants and their incarnations from calls to
`live_devices`. `live_devices` runs a barrier across tasks and returns the set
of currently alive processes along with their incarnations. This CL pipes this
information through the C API, so that open source clients can use the
information. Concretely, the incarnations of every task are passed to an
executable when it is run.
PiperOrigin-RevId: 810549686
PiperOrigin-RevId: 810553684
This dependency is not needed but in the way for ROCm builds because it requires CUDA headers. PiperOrigin-RevId: 810561869
We extend the autotuner to the scaled-dot case. The autotuner has a very restrictive set of config for the scaled dot due to the issues of the scaled-dot -> regular dot AccelerateMatmul rewriter of Triton. The numerics is correct for the config that actually works but requires a transpose for the rhs scales. For some reason the scales should have mk and nk shapes. PiperOrigin-RevId: 810568117
The error: ``` ERROR: .../tensorflow/python/BUILD:1528:15: Mandatory attribute 'actual' resolved to 'None' after evaluating 'select' expression ``` Reverts e1ca152 PiperOrigin-RevId: 810568791
Migrate away from older types. PiperOrigin-RevId: 810574017
PiperOrigin-RevId: 810575107
Enable the StableHLO equivalent of MHLO's similarly named pass, which runs right before lowering to HLO. PiperOrigin-RevId: 810575154
PiperOrigin-RevId: 810589369
…alled CUDA packages. Enable CUDA forward-compatibility mode in all RBE jobs by default. Forward compatibility mode in hermetic CUDA allows the linker to use the user-mode driver from Bazel cache, so there is no need to install UMD in the RBE Docker image. UMD on RBE machines is rarely updated, thus RBE jobs need forward compatibility mode to enable the most recent CUDA features usage in the tests. The non-RBE job runners are updated more often, hence we can update the drivers on those machines and not rely on forward compatibility mode. PiperOrigin-RevId: 810595379
The urls for all python versions can use `pgo+lto-full`. PiperOrigin-RevId: 810595592
…don't divide size. Given the following reshape, * Source: shape 4, sharding [8]<=[8] * Target: shape 2x2 Before this change, we infer that the target sharding is [2,4]<=[8] and convert it into a single reshape without any collective operations, which is wrong. With this change, we infer the sharding to be `[2,2]<=[4]` and generate the correct partitioner result with HALO exchange. PiperOrigin-RevId: 810621300
PiperOrigin-RevId: 810628778
The new emitters don't need this. PiperOrigin-RevId: 810630948
PiperOrigin-RevId: 810633642
This patch, corresponding to Raft PR #2807, updates `select_warpsort.cuh` to use `cub::Traits` for bitwise representation of floating-point and signed integer types, ensuring more stable comparisons, especially for infinities and NaNs. It also includes necessary adjustments to tests. PiperOrigin-RevId: 810634702
A simple class with minimal llvm dependencies capable of loading AOT CPU compilation results. PiperOrigin-RevId: 810642399
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.
12 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 : )