All pull requests

Pull requests

• checkpoint_conversion: materialize lazy tensors on the single-device save path

  #5989·halehri opened Oct 27, 2026·3 of 4 tasks·

  ·
  1

  1

• fix(vllm): use [gate | up] prefused MoE layout under expert parallelism

  #5988·chiajunglien opened Oct 27, 2026Collaborator·4 tasks done·

  ·
  4

• fix(rl): add the local_layers logical-axis rule to rl.yml

  #5987·chiajunglien opened Oct 27, 2026Collaborator·4 tasks done·

  ·
  4

• fix(tests): build attribute-free devices in CountSlicesTest

  #5984·gulsumgudukbay opened Oct 27, 2026Collaborator·4 tasks done·

  ·

• fix(attention): pass the Q-scaling SplashConfig options only when enabled

  #5983·gulsumgudukbay opened Oct 27, 2026Collaborator·4 tasks done·

  ·

• Add DeepSeek-V4.1-Flash text backbone and FP8/FP4 weight dequantization

  gemini-review
  #5977·AusarYao28 opened Oct 26, 2026Collaborator·2/3·4 tasks done·

  ·
  5

• Stream HuggingFace safetensors checkpoints layer-by-layer into HBM

  gemini-review
  #5976·AusarYao28 opened Oct 26, 2026Collaborator·1/3·4 tasks done·

  ·
  6

• Recover elastic training from losing the slice that holds the default device

  #5975·ecnal-cienet opened Oct 26, 2026Collaborator·4 tasks done·

  ·
  4

• Add Treescope guide and demo notebook

  #5974·d0t2 opened Oct 26, 2026Collaborator·4 tasks done·

  ·
  5

• Hide FSDP prefetch backward reduce-scatters by deferring them to the next layer

  pull ready
  #5973·darisoy opened Oct 26, 2026Collaborator·7 tasks done·

  ·
  4

• MoE: native FP8 QAT for routed-expert GMMs (FP8 forward + FP8-code backward dLHS), and fix the dropless full-buffer gather index

  #5969·shuningjin opened Oct 26, 2026Collaborator·4 tasks done·

  ·
  1

• Export Olmo 3 to Hugging Face with the correct shapes and config.json.

  #5967·SUPR1M3 opened Oct 26, 2026·4 tasks done·

  ·
  1

  2

• Release LM head logits cotangents to avoid large memory lifetimes

  pull ready
  #5966·dinodeep opened Oct 25, 2026Collaborator·4 tasks done·

  ·
  17

• Add multi-pass emulation workflow entry points, YAML presets, docs, and scheduled e2e SFT test

  #5964·parsley9877 opened Oct 25, 2026Collaborator·5/6·4 tasks done·

  ·
  4

• Add runtime module/op interception and non-intrusive MaxText/Tunix hooks for multi-pass emulation

  #5963·parsley9877 opened Oct 25, 2026Collaborator·4/6·4 tasks done·

  ·
  10

• Add multi-pass checkpoint materialization, error/SNR inspection, and metadata utilities

  #5962·parsley9877 opened Oct 25, 2026Collaborator·3/6·4 tasks done·

  ·
  10

• Add dtype-agnostic multi-pass residual decomposition and custom-VJP GEMM kernel

  #5960·parsley9877 opened Oct 25, 2026Collaborator·2/6·4 tasks done·

  ·
  9

• Add multi-pass emulation config schema, parametric dtypes, presets, and plugin base

  #5959·parsley9877 opened Oct 25, 2026Collaborator·1/6·4 tasks done·

  ·
  15

• Make HLO/jaxpr dump dir defaults in types.py match base.yml

  #5958·Sonica-B opened Oct 25, 2026·3 of 4 tasks·

  ·
  1

• Opt-in decode embedding lookup in whole 16-row tiles

  #5956·lokic233 opened Oct 24, 2026·4 tasks·

  ·
  2

• Use same model loop unroll context for all gradient accumulation values

  pull ready
  #5954·dinodeep opened Oct 24, 2026Collaborator·4 tasks done·

  ·
  5

• This is a small modification of https://github.com/AI-Hypercomputer/maxtext/pull/5882 by @gobbleturk

  #5952·copybara-service[bot] opened Oct 24, 2026Bot·

  ·
  

• Fix Engram test reference hash mapping bounds to int32

  gemini-review
  #5947·AusarYao28 opened Oct 24, 2026Collaborator·4 tasks done·

  ·
  5

• Lazily import HF model configs so pre-training doesn't load transformers/torchvision/triton at import

  #5944·lydhr opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  5

• Remove transpose_rhs from megablox ops.gmm

  pull ready
  #5943·CaptainO5 opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  7