All pull requests

Pull requests

• Store skip_step_on_spikes flag as int32 for Pathways persistence.

  #6280·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

  

• MoE: add moe_fold_expert_scale_into_wo to apply per_expert_scale to the sharded wo instead of gathering it per token. Default off.

  #6279·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

• fix(testing): restore relative-tolerance loss check for checkpoint_save_restore

  skip-ci-tests
  #6277·ryanliuCienet opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  5

• chore(deps): upgrade qwix to 0.1.9

  #6275·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

  

• Add m3 sharding annotations for Qwen3.8-Flash-Next

  skip-ci-tests
  #6274·tissue3 opened Nov 12, 2026Collaborator·2 of 4 tasks·

  ·
  1

• Add dedicated FSDP prefetch post-scheduler overlap transform

  pull ready
  #6271·darisoy opened Nov 11, 2026Collaborator·7 tasks done·

  ·
  4

• Decouple GMM kernel backend dispatch from quantization logic.

  #6270·copybara-service[bot] opened Nov 11, 2026Bot·4 tasks done·

  ·
  1

  

• Fix packed sequence masking and pre-Hopper fallback in cuDNN JAX attention.

  #6269·copybara-service[bot] opened Nov 11, 2026Bot·

  ·
  1

  

• Split MaxText Docker builds into dependency-only and full images.

  #6268·copybara-service[bot] opened Nov 11, 2026Bot·

  ·
  1

  

• Ensure MaxText HLO dumps upload reliably on failure and cleanup.

  #6267·copybara-service[bot] opened Nov 11, 2026Bot·

  ·
  1

  

• Preserve Qwix quantization across custom_vjp FSDP prefetch backward pass

  #6266·darisoy opened Nov 11, 2026Collaborator·7 tasks done·

  ·
  4

• Prebuild CI dependency images instead of resolving them per test job

  gemini-reviewscheduled-only
  #6265·SurbhiJainUSC opened Nov 11, 2026Collaborator·4 tasks done·

  ·
  10

• Add native-resolution image preprocessing for Gemma-4

  skip-ci-tests
  #6262·nuinashco-ks opened Nov 11, 2026·2 of 4 tasks·

  ·

• DSV4: Add mhc_split_axis_contraction to contract the mHC rate and embed axes separately

  skip-ci-tests
  #6261·RSudarsanan opened Nov 11, 2026Contributor·4 tasks done·

  ·
  3

• mesh: add mesh_ring_axis to give a mesh axis physical rings on a torus without wraparound

  skip-ci-tests
  #6260·RSudarsanan opened Nov 11, 2026Contributor·4 tasks done·

  ·
  3

• Flash attention: replicate the kv head axis for MQA

  skip-ci-tests
  #6258·arminghobadi opened Nov 11, 2026Contributor·4 tasks done·

  ·
  4

• Cast RoutedMoE capacity-factor combine output back to module dtype.

  #6256·copybara-service[bot] opened Nov 11, 2026Bot·

  ·
  1

  1

• [MTC] Add backup=True to save_checkpoint

  #6254·copybara-service[bot] opened Nov 11, 2026Bot·

  ·

• Install CI test dependencies through the Airlock PyPI mirror when reachable.

  #6253·copybara-service[bot] opened Nov 11, 2026Bot·

  ·
  1

  

• Add m3 rematerialization, hybrid loop barriers, and activation savepoints

  skip-ci-tests
  #6249·AusarYao28 opened Nov 10, 2026Collaborator·4 tasks done·

  ·
  8

• Training engine: AOT weight-sync staging estimate, standalone script cleanups, and tests

  #6246·copybara-service[bot] opened Nov 10, 2026Bot·4 tasks done·

  ·
  1

• feat(kimi-k3): add Kimi-K3 model (KDA/MLA hybrid, latent MoE, AttnRes, scan_layers)

  skip-ci-tests
  #6243·snehalv2002 opened Nov 10, 2026Collaborator·4 tasks done·

  ·
  5

• Re-pin Gemma4 RL scripts to cloud-pathways 20261110-jax_0.10.1 digests.

  #6240·copybara-service[bot] opened Nov 10, 2026Bot·

  ·
  1

• Re-pin and bump JAX versions of Pathways image references to cloud-pathways.

  #6239·copybara-service[bot] opened Nov 10, 2026Bot·

  ·
  1

• Use cloud-pathways sidecar image repo in MaxText build script.

  #6238·copybara-service[bot] opened Nov 10, 2026Bot·

  ·
  1