All pull requests

Pull requests

• Add a Qwix multi-pass quantization to MaxText.

  #6185·copybara-service[bot] opened Nov 6, 2026Bot·

  ·
  1

  

• Accept Python int group_offset in ops.gmm and gmm_v2 / tgmm_v2

  skip-ci-tests
  #6184·nigelz-goog opened Nov 6, 2026·4 tasks done·

  ·
  1

• YaRN RoPE: rotate in float32 cos/sin on the default (non-pairwise) path

  skip-ci-tests
  #6183·nigelz-goog opened Nov 6, 2026·4 tasks done·

  ·
  1

• Training engine: parameter freezing and optimizer mask refactoring

  #6181·copybara-service[bot] opened Nov 6, 2026Bot·4 tasks done·

  ·
  1

• Add Muse-Spark-Open-816B-A42B model To Maxtext

  gemini-reviewskip-ci-tests
  #6180·JamesDeng42 opened Nov 6, 2026Collaborator·4 tasks done·

  ·
  8

• extract float router logits across transformers >= 5.22 in granitemoe_test

  skip-ci-tests
  #6179·entrpn opened Nov 6, 2026Collaborator·4 tasks done·

  ·
  3

• feat(kimi-k3): scan_layers=True with scanned checkpoint conversion

  skip-ci-tests
  #6178·snehalv2002 opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• Move slow glm5_3 and deepseek unscanned compile tests from cpu-unit to cpu-integration

  scheduled-onlyskip-ci-tests
  #6177·snehalv2002 opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• Cap SparseCore ragged_gather_reduce_v2 col_chunk_size on large Ring-of-Experts inputs

  skip-ci-tests
  #6175·Rohan-Bierneni opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• Add optimizer NaN/Inf and pre-clipping grad-norm spike guards across pre-train and RL engine

  skip-ci-tests
  #6174·Rohan-Bierneni opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• Add fixed random-logit routing for TE MoE benchmarks

  skip-ci-tests
  #6173·abhinavgoel95 opened Nov 5, 2026Contributor·4 tasks done·

  ·
  5

• feat(configs): add models/deprecated/ fallback resolution and Depreca…

  skip-ci-tests
  #6171·entrpn opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  6

• Merge forward and backward scans in NNX vocabulary tiling loss

  skip-ci-tests
  #6170·Rohan-Bierneni opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• Improve large-scale startup sharding, synthetic data caching, CPU LR logging, and host-offload init

  skip-ci-tests
  #6169·Rohan-Bierneni opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• feat: Onboard GLM-5.2 and GLM-5.1 with native IndexShare

  skip-ci-tests
  #6168·h1t35h opened Nov 5, 2026Contributor·4 tasks done·

  ·
  5

• Fix PyPI and OS vulnerabilities in MaxText Docker images

  skip-ci-tests
  #6167·KevinYenky opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  9

• Label MaxText RL ML Diagnostics runs with rl_orchestrator=tunix.

  #6166·copybara-service[bot] opened Nov 5, 2026Bot·

  ·
  1

• Bring up Wan 2.2 TI2V 5B text-to-video in Maxtext-M3

  skip-ci-tests
  #6165·lydhr opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• Add shared m3 MoE executor with EP/TP, grouped GMM, and wgrad overlap

  skip-ci-tests
  #6164·AusarYao28 opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• Add shared m3 Splash attention kernels with SEQ_MINOR, GQA, and SWA

  skip-ci-tests
  #6163·AusarYao28 opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  10

• Fix Llama 4 decode: chunked attention picks keys by KV-cache slot instead of token position

  skip-ci-tests
  #6162·EvanZhuang opened Nov 5, 2026·4 tasks done·

  ·
  1

• Read colocated ArrayRecord file metadata once on the controller.

  #6161·copybara-service[bot] opened Nov 5, 2026Bot·

  ·
  1

  

• Fix Llama 4 Scout RoPE: llama4-17b-16e runs with Llama 3.1's scaling (8/1/4) instead of Scout's 16/1/1

  skip-ci-tests
  #6160·EvanZhuang opened Nov 5, 2026·4 tasks done·

  ·
  1

• Fix Gemma 3 to_huggingface export: global layers get rope_theta 10,000 instead of 1,000,000

  skip-ci-tests
  #6159·EvanZhuang opened Nov 5, 2026·4 tasks done·

  ·
  1

• Fix llama3.3-70b RoPE: Meta-converted checkpoints run with half-split pairs and no llama3 scaling

  skip-ci-tests
  #6158·EvanZhuang opened Nov 5, 2026·4 tasks done·

  ·
  1