Skip to content

[pull] master from tensorflow:master#1974

Merged
pull[bot] merged 15 commits into
GesuBackups:masterfrom
tensorflow:master
Jun 27, 2026
Merged

[pull] master from tensorflow:master#1974
pull[bot] merged 15 commits into
GesuBackups:masterfrom
tensorflow:master

Conversation

@pull

@pull pull Bot commented Jun 27, 2026

Copy link
Copy Markdown

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 : )

akuegel and others added 15 commits June 26, 2026 12:31
The xla_framework dialect and its associated passes
(LegalizeXLAFrameworkToLLVM, OutlineWithXLAFramework) are legacy and no longer
used. Delete them and their tests, and remove their registration from various
opt tools.

PiperOrigin-RevId: 938712305
Initial XLA PR to support unstable TopK operations in JAX/XLA.

By introducing the is_stable flag (defaulting to true for backward compatibility), we can eventually expose is_stable=False at the Python JAX level and dispatch to NVIDIA RAFT select_k kernel

Changes:
- Added the is_stable attr to CHLO_TopKOp and MHLO_TopKOp
- Updated CHLO to MHLO and StableHLO legalization patterns to propagate the flag
- Plumbed the attribute through HloTopKInstruction ToProto and hlo parser.
- Updated IdenticalSlowPath to account for the flag.
- Modified topk_rewriter.cc to ensure that when TopK is lowered or decomposed, is_stable is preserved on the resulting Sort instruction.
- Updated CustomCall TopK generator to print backend config (e.g. {is_stable = false})
PiperOrigin-RevId: 938761815
…or AssignBuffersForComputations, similar to go/xla-fast-buffer-assignment. It adds `FAST_MERGE` to `AssignmentAlgorithmForComputationsWithoutOrderingProto` and implements `FastMergeBufferAllocationsManagerForComputationsWithoutOrdering` to track buffer liveness and reuse expired allocations from a free pool.

Instead of the default O(N * M) naive allocation scan (where N is the number of buffers and M is the number of allocations, leading to O(N^2) worst-case), this algorithm achieves O(N log M) time complexity for allocation reuse.

Key features:

1. Incoming buffers are processed in increasing order of their live range start times (BufferOrder::kLiveRangeStart). It tracks active allocations in a min-heap (`active_allocations_`) ordered by their maximum live range end time.

2. Before assigning a new buffer, it advances the sweep-line (`UpdateLiveness`), moving any active allocations that have fully expired (max_end_time <= new_buffer.min_start) into a free pool. By this, allocations in the free pool are guaranteed to have no live range conflicts with the incoming buffer.

3. The free pool (`free_pool_`) indexes available allocations by color and size using balanced trees (absl::btree_multimap). Finding the smallest fitting reusable allocation takes O(log M) time via `lower_bound`, avoiding linear scans over all existing allocations.

PiperOrigin-RevId: 938794789
…HLO <-> StableHLO translation.

Custom-call operand/result layout constraints that use a non-default memory
space are now preserved across the HLO <-> StableHLO round trip via the
mhlo.operand_memory_spaces / mhlo.result_memory_spaces attributes, instead of
failing with UNIMPLEMENTED.

PiperOrigin-RevId: 938812415
The error message incorrectly referred to "consumers" and used `users().size()` instead of "operands" and `operand_count()`.

PiperOrigin-RevId: 938818957
…re strictly non-negative.

PiperOrigin-RevId: 938844012
… found before calling RemoveAt in the Remove utility. Added unit test.

PiperOrigin-RevId: 938845085
Delete legacy TPU compile/execute ops, compilation cache, and related C APIs.

PiperOrigin-RevId: 938847080
…dim is

divisible by 8 / bits_per_element.

PiperOrigin-RevId: 938847861
… TFLite concatenation kernel.

PiperOrigin-RevId: 938856354
@pull pull Bot locked and limited conversation to collaborators Jun 27, 2026
@pull pull Bot added the ⤵️ pull label Jun 27, 2026
@pull pull Bot merged commit a112d5a into GesuBackups:master Jun 27, 2026
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

9 participants