All pull requests

Pull requests

• Document how to restore a Multi-Tier Checkpointing job from a specific backup.

  #5683·copybara-service[bot] opened Oct 13, 2026Bot·

  ·

• MaxText: add mtp_norm_main_hidden_state to feed the final-norm output into MTP.

  #5681·copybara-service[bot] opened Oct 13, 2026Bot·

  ·
  1

• Add DeepSeek-V4-Pro (deepseek4-1600b) model, conversion and e2e tests (b/568401715)

  gemini-review
  #5679·AusarYao28 opened Oct 12, 2026Collaborator·3/3·4 tasks done·

  ·
  6

• Honor remat_policy in DeepSeek-V4 train-mode layers (b/568401793)

  gemini-review
  #5678·AusarYao28 opened Oct 12, 2026Collaborator·2/3·4 tasks done·

  ·
  9

• Fix DeepSeek-V4 decode after padded prefill and sliding-window decode masks (b/572863708)

  gemini-review
  #5677·AusarYao28 opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  3

• Fix DeepSeek router numerics: shared-expert clamp, fp32 router bias (b/566424998)

  gemini-review
  #5676·AusarYao28 opened Oct 12, 2026Collaborator·1/3·4 tasks done·

  ·
  4

• Run TensorFlow/JetStream-dependent tests in PR CI

  #5675·snehalv2002 opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  4

• Remove pallas_tpu_ops dependency from MaxText.

  #5674·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Training engine: MoE drop audit (log_moe_drops) and final micro-batch dropless program

  #5673·ultrons opened Oct 12, 2026Collaborator·

  ·
  5

• MoE: exclude padding tokens from routing (moe_exclude_padding_from_routing)

  #5672·ultrons opened Oct 12, 2026Collaborator·

  ·
  3

• Add Qwen3.5-9B multimodal support

  #5669·hengtaoguo opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  2

• Register OLMo3 shapes for Hugging Face checkpoint export

  #5668·jabrailkhalil opened Oct 12, 2026·3 of 4 tasks·

  ·
  3

• Optimize GDN triangular inverse with xor_dnc and support gdn_chunk_size=128

  #5667·Rohan-Bierneni opened Oct 12, 2026Collaborator·

  ·
  2

• Fix sub-block KDA inverse miscompile on TPU v4

  #5665·abhishekraok opened Oct 12, 2026·

  ·
  1

• Pass MaxText application framework label to managed ML Diagnostics runs.

  #5663·copybara-service[bot] opened Oct 12, 2026Bot·

  ·
  1

• Splash attention: support FP8 LOG2E Q prescaling and dkv scaled-Q reuse

  #5661·shuningjin opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  4

• feat(ci): Gate TPU/GPU CI suites on PRs behind the ready_for_full_ci label

  ready_for_full_ci
  #5660·mesakhcienet opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  3

• feat(goodput): use get_goodput_job_name in setup_checkpoint_logger

  #5659·KevinYenky opened Oct 12, 2026Collaborator·4 tasks done·

  ·
  4

• Compute YaRN rotary frequencies inline without 2D gather lookup.

  #5657·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Keep gathers on SparseCore for TPU7x CI; re-enable test_gmm_grad_equivalence on TPU7x

  #5655·shuningjin opened Oct 11, 2026Collaborator·2 of 4 tasks·

  ·

• Tokamax Splash Attention: expose sa_dkv_diag_skip and sa_dkv_diag_skip_dq

  #5652·RissyRan opened Oct 11, 2026Collaborator·

  ·
  3

• Vocab tiling: add vocab_tiling_seq_major to tile the loss along the sequence

  #5651·RissyRan opened Oct 11, 2026Collaborator·

  ·
  3

• Document shard_mode performance and the explicit-sharding tuning flags

  #5650·NuojCheng opened Oct 11, 2026Collaborator·3 of 4 tasks·

  ·

• Contract the tied LM head over the embedding table's feature axis

  #5649·NuojCheng opened Oct 11, 2026Collaborator·4 tasks done·

  ·
  2

• Build YaRN cis values from cos and sin instead of complex exp

  #5648·NuojCheng opened Oct 11, 2026Collaborator·4 tasks done·

  ·