Default Branch

4f8e6d470a · Fix BUILD dependencies in xla/python · Updated 2026-10-25 02:11:06 +08:00

Branches

4eeb4a35f7 · Constant-fold RESIZE_BILINEAR and RESIZE_NEAREST_NEIGHBOR. · Updated 2026-10-22 03:52:43 +08:00

346
1

14bea607cd · Constant-fold BROADCAST_TO. · Updated 2026-10-22 03:47:37 +08:00

346
1

046a582230 · Fix BUILD dependencies in xla/tools · Updated 2026-10-22 02:15:39 +08:00

351
1

d5f9e6411e · Automated Code Change · Updated 2026-10-21 22:00:30 +08:00

364
1

1fe1e69597 · Automated Code Change · Updated 2026-10-21 19:14:38 +08:00

377
1

0302320690 · Automated Code Change · Updated 2026-10-21 19:12:03 +08:00

377
1

6739103801 · Automated Code Change · Updated 2026-10-21 19:09:16 +08:00

377
1

d6c3cd61a1 · Automated Code Change · Updated 2026-10-21 19:03:38 +08:00

377
1

6aeed2f7c6 · Automated Code Change · Updated 2026-10-21 17:34:15 +08:00

382
1

bf412d5b72 · Automated Code Change · Updated 2026-10-21 16:26:57 +08:00

387
1

a65270a1f5 · [PJRT] Introduce PjRtBuildStamp value type to provide an immutable representation of compiler build provenance, binary hashes, compile-time components, and extensible attributes. · Updated 2026-10-21 15:34:01 +08:00

389
1

454d5413b7 · Automated Code Change · Updated 2026-10-21 15:10:30 +08:00

391
1

818b4d0998 · [TF2XLA][SDY] Only export (don't propagate) Shardy shardings in XlaCallModuleLoader when the outer TF compiler uses GSPMD. · Updated 2026-10-21 14:19:14 +08:00

395
1

bf4d395de1 · Automated Code Change · Updated 2026-10-21 13:26:36 +08:00

397
1

069d4a381f · Bump pillow from 12.3.0 to 13.0.0 · Updated 2026-10-21 07:01:17 +08:00

422
1

999b09b867 · Update TensorFlow Java documentation link in tensorflow/java/README.md to https://tensorflow.github.io/java/install. · Updated 2026-10-21 06:22:13 +08:00

425
1

7f49aedf32 · Use rules_ml_toolchain for macOS aarch64 compilation · Updated 2026-10-21 06:14:41 +08:00

426
1

b99e39825b · In AlgebraicSimplifierVisitor::RemoveTransposesFromDotOperands, rewriting dot(transpose(a, P), transpose(b, P), dnums) into transpose(dot(a, b, dnums), P) previously preserved the original dnums and reused P for both the outer transpose and ShapeUtil::PermuteDimensions(P, dot->shape()). When dnums.lhs_batch_dimensions() is not the identity permutation [0, ..., num_batch_dims - 1] and does not commute with P, transpose(dot(a, b, dnums), P) pairs and outputs the wrong batch slices (P[B[P[i]]] instead of P[B[i]]), and applying P instead of InversePermutation to dot->shape() computes the wrong intermediate shape when the composed batch permutation has a cycle of length >= 3. Canonicalize new_dnums batch dimensions to 0..num_batch_dims-1, compose outer_permutation[i] = permutation[dnums.lhs_batch_dimensions(i)] for batch dimensions, and permute dot->shape() by InversePermutation(outer_permutation). · Updated 2026-10-21 04:10:03 +08:00

433
1

b58b221889 · Release the Python GIL in TF_ListPhysicalDevices, TF_ListPluggablePhysicalDevices, TF_GetDeviceDetails, and TFE_NewContext in tfe_wrapper.cc so synchronous C++ device discovery and context creation do not block concurrent Python threads. · Updated 2026-10-20 20:33:39 +08:00

461
1

a5a3908a64 · Skip unused CallGraph and replication analyses in SPMD ArCrsCombiner · Updated 2026-10-20 19:44:23 +08:00

465
1