All pull requests

Pull requests

• Ignore the per-host default JAX cache in multi-process GPU runs

  skip-ci-tests
  #6315·KevinYenky opened Nov 13, 2026Collaborator·4 tasks done·

  ·
  4

• Add BOS to text decode prompts when the tokenizer has a chat template

  #6314·KevinYenky opened Nov 13, 2026Collaborator·4 tasks done·

  ·
  4

• [DeepSeek-V4.1-Flash][Stacked 5/5 on #6311] Full-model v5p-8 layerwise verification harness

  gemini-review
  #6312·AusarYao28 opened Nov 13, 2026Collaborator·4 tasks done·

  ·
  4

• [DeepSeek-V4.1-Flash][Stacked 4/5 on #5997] Multimodal E2E verification suite, golden tools, and AOT gates

  gemini-review
  #6311·AusarYao28 opened Nov 13, 2026Collaborator·4 tasks done·

  ·
  3

• Extend Tokamax Splash attention with fp8 matmuls, VMEM GQA dkv, and fp8 context offload.

  #6310·copybara-service[bot] opened Nov 13, 2026Bot·

  ·
  1

  

• Fuse the Qwen3 MoE expert path into Lineage GMMv2 kernels with in-kernel gather and TC combine.

  #6309·copybara-service[bot] opened Nov 13, 2026Bot·

  ·
  

• Overlap exposed collectives and host prefetch for Qwen3-235B on TPU v7x.

  #6308·copybara-service[bot] opened Nov 13, 2026Bot·

  ·
  

• Wire deferred norm reductions and FSDP prefetch into Qwen3 and NNXDecoder.

  #6307·copybara-service[bot] opened Nov 13, 2026Bot·

  ·
  

• Add fused fp8 QKV/out projection kernels with QK-RMSNorm and RoPE for Qwen3 attention.

  #6305·copybara-service[bot] opened Nov 13, 2026Bot·

  ·
  

• chore(deps): upgrade qwix to 0.1.10

  pull readyskip-ci-tests
  #6304·darisoy opened Nov 13, 2026Collaborator·8 tasks done·

  ·
  4

• [MoE replication 4/4] Wire online hot-expert replication into the ragged all-to-all MoE path

  skip-ci-tests
  #6302·Shuwen-Fang opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  1

• [MoE replication 3/4] Add replica weight-move ops for online hot-expert replication

  skip-ci-tests
  #6301·Shuwen-Fang opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  1

• [MoE replication 2/4] Add replica planners for online hot-expert replication

  skip-ci-tests
  #6300·Shuwen-Fang opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  1

• Add log_ra2a_required_rbf debug flag for ragged all-to-all MoE

  skip-ci-tests
  #6299·Shuwen-Fang opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  6

• Add synthetic MoE routing benchmark harness under benchmarks/moe_synthetic_routing

  skip-ci-tests
  #6298·Shuwen-Fang opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  1

• Remove legacy XPK wrapper and XPK instructions for OLMo-3 7B

  skip-ci-tests
  #6297·cj-klein opened Nov 12, 2026·

  ·
  1

• Support dynamic scaling for moe_quantize_token_all_gather.

  #6296·copybara-service[bot] opened Nov 12, 2026Bot·4 tasks done·

  ·
  1

  

• Raise instead of assuming 1 device when config validation can't get the device count

  pull ready
  #6291·NuojCheng opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  1

  3

• Internal change

  #6290·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

  

• Support full attention on TPU Ulysses CP in MaxText.

  #6289·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

  

• Support dynamic (absmax) weight scaling for MoE QWAG

  #6288·copybara-service[bot] opened Nov 12, 2026Bot·4 tasks done·

  ·
  1

  

• MoE: add moe_wi_dlhs_merged_k to compute the routed wi_0/wi_1 input gradient with one gmm_v2 DLHS. Default off.

  #6287·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

• Adapt the experimental Lineage DeepSeek-V3 671B (+ 1 MTP layer) trainer to run on a single 64-chip TPU v7x slice (tpu7x:4x4x4, 128 TensorCores) and add tuned 8K and 16K long-context recipes under src/maxtext/experimental/lineage/recipes/.

  #6286·copybara-service[bot] opened Nov 12, 2026Bot·

  ·
  1

  

• Expand MaxText m3 conformance suite with Checks 9a–9c and 10–12

  #6285·copybara-service[bot] opened Nov 12, 2026Bot·4 tasks done·

  ·
  1

  

• Add tiled block-sparse FlashAttention with 32-bit packed bitmap mask for m3 Qwen3.8 QSA

  gemini-review
  #6283·AusarYao28 opened Nov 12, 2026Collaborator·4 tasks done·

  ·
  15