All pull requests

Pull requests

• Gemma4-26b v7x recipe: enable xla_tpu_use_minor_sharding_for_major_trivial_input (+3.3%).

  #6125·copybara-service[bot] opened Nov 4, 2026Bot·

  ·
  1

• MoE: add moe_weights_on_activation and moe_sc_gather_reduce (SparseCore gather-reduce for the top-k sums). Default off.

  #6124·copybara-service[bot] opened Nov 4, 2026Bot·

  ·

• Gemma4-26b v7x recipe: Tokamax GMM v2 with Gemma4-shaped tiles (+7.0%).

  #6123·copybara-service[bot] opened Nov 4, 2026Bot·

  ·
  1

• Original PR authors: Igor Tsvetkov (igorts), Wenxin Dong (wenxindongwork), Juan Acevedo (jfacevedo).

  #6122·copybara-service[bot] opened Nov 4, 2026Bot·4 tasks done·

  ·
  1

• Add a Gemma4-26b pre-training recipe for TPU v7x (configs/tpu/tpu7x/gemma4_26b.sh).

  #6121·copybara-service[bot] opened Nov 4, 2026Bot·

  ·
  1

• Upgrade JAX to 0.12.0.1 for Pre-Training

  scheduled-only
  #6120·SurbhiJainUSC opened Nov 4, 2026Collaborator·4 tasks done·

  ·
  1

• Update default Pathways container image paths in MaxText benchmarks to Louhi registry.

  #6119·copybara-service[bot] opened Nov 4, 2026Bot·

  ·
  1

• Port FP8 MoE fake quantization and weight sync from atwigg/mlperf

  skip-ci-tests
  #6118·igorts-git opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  10

• Added a priority scheduling logic and control knob to the gpu custom scheduling pass

  skip-ci-tests
  #6117·Tixxx opened Nov 3, 2026Contributor·

  ·
  5

• Share the single-axis stacking rule between the safetensors_dynamic loader and the converters

  pull ready
  #6116·ecnal-cienet opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• Fix the pipeline explicit-sharding tests on tpu7x

  pull readyscheduled-only
  #6115·NuojCheng opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• MoE: use gmm_v2 as the only TPU grouped matmul; deprecate megablox and tokamax gmm

  #6114·d0t2 opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• (snapshotting fix) Limit GCS glob listings to the pattern's literal prefix

  skip-ci-tests
  #6113·ecnal-cienet opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• Add model-agnostic m3 CPU conformance spec, single-device training-engine Checks 1-4, and unit tests.

  #6112·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

  

• feat(m3): add remat policies (m3/infra/remat.py) and the WanModel remat hook

  skip-ci-tests
  #6111·lydhr opened Nov 3, 2026Collaborator·4 tasks·

  ·
  1

• Mask left-padding tokens in Qwen3NextGatedDeltaNet and guard l2norm on zero vectors

  skip-ci-tests
  #6110·niting opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• Re-enable a Google-internal TPU test. BUILD-only change.

  #6109·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

  

• test(m3): move diffuser scheduler unit test to tests/unit/m3/

  skip-ci-tests
  #6108·lydhr opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  4

• Overview

  #6107·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Overview

  #6106·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• feat: unified parameter freezing across differentiation, grad norm, and optimizer

  gemini-reviewskip-ci-tests
  #6105·weikuo0506 opened Nov 3, 2026Collaborator·4 tasks done·

  ·
  7

• Overview

  #6103·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Overview

  #6102·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Add moe_init_under_default_expert_sharding: draw routed expert kernels under the default expert sharding so a different storage layout (e.g. shard_exp_on_fsdp) keeps initial weights identical. Default off.

  #6098·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

• Gemma4: add the custom-remat tensor moe_routed_output (routed MoE output after unsort and top-k combine). Default remat; device saves it so the backward skips replaying the wo GMM, unsort and combine.

  #6099·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1