All pull requests

Pull requests

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

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

  ·
  1

• Update README to put versioning notes below MaxText description

  #5918·bvandermoon opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  3

• Fix DeepSeek-V4 remat initialization with Qwix FP8

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

  ·
  1

  1

• Run tests/dev in ci_pipeline.yml as a standalone Development Tests job

  #5916·SurbhiJainUSC opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  3

• Replace scatter with one-hot reduction in RoutedMoE.dense_matmul.

  #5915·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• [Pallas:SC] Support non-power-of-2 group sizes via unsigned division in ragged gather-reduce v2

  #5913·denis-mil opened Oct 22, 2026·4 tasks done·

  ·
  1

• Update FSDP weight prefetching guide with inhomogeneous scans, EP, and recommended flags

  #5912·darisoy opened Oct 22, 2026Collaborator·7 tasks done·

  ·
  4

• Add Llama 3.1 8B GPU E2E test scripts (to_mt, 1node, 2node, to_hf)

  #5908·KevinYenky opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  6

• Support multi-arch (linux/arm64 and linux/amd64) GPU Docker image builds

  #5907·KevinYenky opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  4

• Add Ming checkpoint conversion and reproducible validation tools

  #5904·hengtaoguo opened Oct 22, 2026Collaborator·3/3·3 of 4 tasks·

  ·
  4

• Add Ming Whisper audio input and inference preprocessing

  #5903·hengtaoguo opened Oct 22, 2026Collaborator·2/3·3 of 4 tasks·

  ·
  7

• Add Ming hybrid KDA/MLA text decoder with scanned execution

  gemini-review
  #5902·hengtaoguo opened Oct 22, 2026Collaborator·1/3·4 tasks done·

  ·
  20

• Avoid per-step NNX graph traversal in the m3 training loop

  #5900·bvandermoon opened Oct 22, 2026Collaborator·4 tasks done·

  ·
  3

• Add glm5.3-flash-mini, a full-width 5-layer twin of GLM-5.3-Flash

  #5899·RissyRan opened Oct 22, 2026Collaborator·3 of 4 tasks·

  ·
  3

• Restore hierarchical jax.named_scope for modules in XProf traces

  #5895·bvandermoon opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  6

• Avoid MoE weight relayouts with FSDP prefetch via moe_flat_fsdp_weights

  #5890·darisoy opened Oct 21, 2026Collaborator·7 tasks done·

  ·
  6

• Optimize CI worker sizing, baseline pruning, and test suite scoping

  #5889·SurbhiJainUSC opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  1

• Multi-device FSDP compilation for Qwen3.8-Flash-Next previously ran out of HBM on 128 TPU v5p chips (vf=4x4x8) due to missing activation checkpointing inside the scanned decoder cycle, batch-dimension replication after the vocab-sharded 51B PLE n-gram gather, and global token-expert sorting across all chips in Qwen3_8Experts.

  #5888·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Fix DPO/ORPO evaluation loss averaging and early stopping

  #5886·barnes-vidit opened Oct 21, 2026Contributor·4 tasks done·

  ·
  3

• [MLPerf RL 397B] Final 4,589 tok/s/chip trainer kernel & config stack (supersedes #5799)

  #5885·uwlei opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  15

• Add DeepSeek-V4 chat encoding to SFT preprocessing

  gemini-reviewpull ready
  #5883·AusarYao28 opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  31

• Add performance_reviewer skill and integrate into Gemini PR review

  gemini-review
  #5882·gobbleturk opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  6

• [maxtext:attention_op] Support non-causal, sliding window and soft-cap in the GPU Tokamax (ragged) attention path.

  #5881·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  1

• OLMoE3: log the OLMo-core hero's metrics, fix z-loss under GA, and add OLMo-core's perplexity evals

  #5880·AkshitaB opened Oct 21, 2026·

  ·
  5

• feat(benchmarks): expose elastic training, goodput, checkpointing and data-input overrides in recipes and CTK runner

  #5879·hsuan-lun-chiang opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  5