All pull requests

Pull requests

• [MLPerf][Serving] Enable DCP attention routing and is_decode propagation for MaxText vLLM serving (#5621)

  #5634·wenxindongwork opened Oct 10, 2026Collaborator·

  ·
  3

• Migrate DeepSeek e2e tests to to_maxtext and fix scanned MoE gate stacking & Grain iterator cleanup

  #5633·SurbhiJainUSC opened Oct 9, 2026Collaborator·4 tasks done·

  ·
  4

• Fix(Checkpoint): Compute FP8 MoE scales and quantize weights when loading unquantized checkpoint into FP8 model

  #5629·wenxindongwork opened Oct 9, 2026Collaborator·

  ·
  2

• [GDN] Layout fixes for the Pallas kernel path: bf16 qkv input, natural-layout backward kernel, head-major backward glue (opt-in)

  #5628·tissue3 opened Oct 9, 2026Collaborator·4 tasks done·

  ·
  7

• [GDN] Sequence-CP two-pass forward: pass-1 states-only, in-kernel M_local, pass-2 t_inv reuse (opt-in)

  #5627·tissue3 opened Oct 9, 2026Collaborator·4 tasks done·

  ·
  4

• [MLPerf][MoE] Forward ragged_gather_reduce_version to fused_moe_func in vllm_rpa

  #5620·wenxindongwork opened Oct 8, 2026Collaborator·

  ·
  2

• MoE: invert sort permutations with an O(N) scatter

  gemini-review
  #5617·systalyze-ai opened Oct 8, 2026·4 tasks done·

  ·
  5

• [DO NOT REVIEW] Release pipeline v2

  #5616·SurbhiJainUSC opened Oct 8, 2026Collaborator·4 tasks·

  ·
  2

• [DO NOT REVIEW] Add a Github workflow that builds release candidate Docker images weekly

  #5615·SurbhiJainUSC opened Oct 8, 2026Collaborator·4 tasks·

  ·
  2

• Optimize OLMo 3.5 Small (olmoe3-3p5b) training on TPU.

  #5614·copybara-service[bot] opened Oct 8, 2026Bot·

  ·
  

• fix(load_dynamic): Correct weight dtype and -Instruct lookup

  #5610·ecnal-cienet opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  1

• feat(vllm): support Decode Context Parallelism (DCP) for Qwen3.5 in MaxTextForCausalLM

  #5609·sierraisland opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  1

• Attention: per-layer local_context_parallel_strategy and halo KV exchange for LOCAL_SLIDING

  #5608·csgoogle opened Oct 8, 2026Contributor·

  ·
  10

• feat(attention): support vLLM FP8 KV cache in RPA serving and add Trainer FP8 KV fake-quant (fp8_kv_fake_quant)

  #5606·pv97 opened Oct 8, 2026·

  ·
  7

• ci(moss): (4/4) add staging-to-production Exit Gate promotion

  #5604·chiajunglien opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  2

• ci(moss): (2/4) add PyPI SLSA provenance and SBOM attestations

  pull ready
  #5602·chiajunglien opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  2

• ci(moss): (1/4) add SPDX SBOM generation and Syft setup

  #5601·chiajunglien opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  3

• Use real next-token labels at chunk boundaries in C4 MLPerf stream chunking

  #5600·snehalv2002 opened Oct 8, 2026Collaborator·3 of 4 tasks·

  ·
  3

• feat: Add CI metrics collector and automated dashboard pipeline

  gemini-review
  #5599·mesakhcienet opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  6

• Publish Read the Docs versioned documentation only after PyPI release

  #5598·SurbhiJainUSC opened Oct 8, 2026Collaborator·4 tasks done·

  ·
  4

• HF streaming (5/6): Stream HF checkpoints in from_pretrained with safetensors_dynamic

  #5594·snehalv2002 opened Oct 7, 2026Collaborator·5/6·4 tasks·

  ·
  4

• HF streaming (1/6): Make Qwen3.5 HF->MaxText hooks work on several hosts

  #5593·snehalv2002 opened Oct 7, 2026Collaborator·1/6·4 tasks·

  ·
  3

• HF streaming (2/6): Add a SafetensorsReader that reads HF tensors straight into sharded jax.Arrays

  #5592·snehalv2002 opened Oct 7, 2026Collaborator·2/6·4 tasks·

  ·
  4

• HF streaming (6/6): Support DeepSeek-V4 checkpoints in the streaming loader

  #5591·snehalv2002 opened Oct 7, 2026Collaborator·6/6·4 tasks·

  ·
  4

• HF streaming (3/6): Stream safetensors_dynamic checkpoints into MaxText a few layers at a time

  #5590·snehalv2002 opened Oct 7, 2026Collaborator·3/6·4 tasks·

  ·
  4