All pull requests

Pull requests

• Fix DeepSeek-V4 E2E golden config and remove legacy transformers pin

  #5957·AusarYao28 opened Oct 25, 2026Collaborator·4 tasks done·

  ·
  5

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

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

  ·
  2

• Remove logits from cross-entropy residuals to save memory

  #5955·dinodeep opened Oct 24, 2026Collaborator·4 tasks done·

  ·
  7

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

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

  ·
  4

• Reshard Zero-1 train state on device under Pathways single controller.

  #5953·copybara-service[bot] opened Oct 24, 2026Bot·4 tasks done·

  ·
  1

  

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

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

  ·
  

• Write the colocated data iterator state after Orbax creates the checkpoint directory.

  #5951·copybara-service[bot] opened Oct 24, 2026Bot·

  ·
  1

  

• Back up the step before an elastic scale-up interrupts the attempt.

  #5950·copybara-service[bot] opened Oct 24, 2026Bot·

  ·
  1

  

• Shard colocated elastic data across the hosts of the current mesh.

  #5949·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

• Add m3 section to CODEOWNERS

  pull ready
  #5946·bvandermoon opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  4

  

• 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

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

  ·
  4

• Checkpointing: explicit checkpoint_dir and local (GCSFuse) paths under colocated_python

  #5942·igorts-git opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  2

• OLMo 3.5 Small (Option A): 34.26% MFU on 3,072 TPU v4 Chips (8x8x48) with Pathways & McJAX

  #5941·copybara-service[bot] opened Oct 23, 2026Bot·4 tasks done·

  ·
  1

  

• hengtaoguo-m3adapt

  #5938·hengtaoguo opened Oct 23, 2026Collaborator·4 tasks·

  ·
  2

• Fix DeepSeek-V4.1-Flash hc_post precision, quant, and training wiring

  gemini-review
  #5937·AusarYao28 opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  5

• Model lifecycle phase1 poc

  #5935·entrpn opened Oct 23, 2026Collaborator·4 tasks·

  ·
  5

• fix(rl): accept scanned sources in the direct-sync layer coverage check

  #5932·chiajunglien opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  4

• Support pipeline parallelism under shard_mode=explicit

  #5931·NuojCheng opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  8

• Add scan support for m3 Qwen3

  gemini-review
  #5929·hengtaoguo opened Oct 23, 2026Collaborator·4 tasks done·

  ·
  18

• Restart distillation evaluation data for each pass

  #5927·barnes-vidit opened Oct 23, 2026Contributor·4 tasks done·

  ·
  3

• Eliminate suboptimal fusion layout in embeddings.py.

  #5921·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  1

• Support moe_dropless_fallback=layer on the ragged all-to-all MoE path

  #5920·d0t2 opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  11

• Fix DeepSeek-V4 remat initialization with Qwix FP8

  #5917·janbernloehr opened Oct 22, 2026·3 of 4 tasks·

  ·
  1

  1