All pull requests

Pull requests

• [pallas:mosaic] Migrated all uses of pltpu.emit_pipeline* to explicit indices

  #6036·copybara-service[bot] opened Oct 29, 2026Bot·

  ·

• fix(ckpt-conversion): port llama_mistral_mixtral_orbax_to_hf to the NNX training state

  skip-ci-tests
  #6035·ryanliuCienet opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• fix(multimodal): prepend BOS to Gemma 3 multimodal prompts

  skip-ci-tests
  #6034·chiajunglien opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• Add Qwix PyTree/Orbax checkpoint quantization utility and end-to-end workflow tests

  skip-ci-tests
  #6033·parsley9877 opened Oct 29, 2026Collaborator·4 tasks done·

  ·

• Wire Qwix step_context, scan_index_context, and role-scoped model creation

  skip-ci-tests
  #6032·parsley9877 opened Oct 29, 2026Collaborator·4 tasks done·

  ·

• Extend Qwix quantization provider, multipass config, rule scoping, and megablox GMM

  skip-ci-tests
  #6031·parsley9877 opened Oct 29, 2026Collaborator·4 tasks done·

  ·

• ci: schedule the nightly and RC pipelines for the E2E TPU quota window

  skip-ci-tests
  #6030·chiajunglien opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• build(deps): move diffusers to shared base requirements

  skip-ci-tests
  #6029·lydhr opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• fix(vllm): support GPT-OSS on the fused MoE path of MaxTextForCausalLM

  skip-ci-tests
  #6028·chiajunglien opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  6

• fix(gpt-oss): give each scanned GPT-OSS layer its own vLLM KV cache

  skip-ci-tests
  #6027·chiajunglien opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• fix(checkpointing): set the unsafe_rbg PRNG impl in generate_param_only_checkpoint

  skip-ci-tests
  #6026·ryanliuCienet opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• fix(testing): fall back to the MaxText tokenizer in the logit checker's golden mode

  skip-ci-tests
  #6025·ryanliuCienet opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  8

• fix(testing): run the gemma2-2b and qwen3-4b logit checks with dot_product attention in float32

  skip-ci-tests
  #6024·ryanliuCienet opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  4

• Skip TE mHC GPU integration test when Transformer Engine cannot load.

  #6022·copybara-service[bot] opened Oct 29, 2026Bot·

  ·
  1

  

• Training engine: separate accumulate kernel, accumulation dtype, optimizer offload, frozen params and step guards

  skip-ci-tests
  #6020·igorts-git opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  1

• Training engine: harden checkpoint save/restore

  skip-ci-tests
  #6019·igorts-git opened Oct 28, 2026Collaborator·4 tasks done·

  ·

• Training engine: buffer metrics on the host

  skip-ci-tests
  #6017·igorts-git opened Oct 28, 2026Collaborator·4 tasks done·

  ·
  1

• Keep the global batch size constant across elastic slice resizes.

  #6016·copybara-service[bot] opened Oct 28, 2026Bot·

  ·
  1

  

• feat(kimi-k3): optional tokamax KDA kernel (kda_kernel=tokamax)

  skip-ci-tests
  #6014·snehalv2002 opened Oct 28, 2026Collaborator·

  ·

• Add an agent skill to facilitate usage of Goodput measurement

  skip-ci-tests
  #6013·lydhr opened Oct 28, 2026Collaborator·4 tasks done·

  ·
  5

• [Pallas:SC] use get_tpu_info instead of hardcoded values in ragged_gather kernels

  skip-ci-tests
  #6012·denis-mil opened Oct 28, 2026·3 tasks done·

  ·
  1

• Fix MoE ragged_all_to_all dispatch buffer zero-init and local_permute tail sentinel

  gemini-reviewskip-ci-tests
  #6011·AusarYao28 opened Oct 28, 2026Collaborator·4 tasks done·

  ·
  5

• Forward speculative decoding metadata in Qwen3NextGatedDeltaNet and RPA serving attention

  #6010·pv97 opened Oct 28, 2026·

  ·
  1

• Apply YaRN to Olmo 3 full-attention layers only.

  #6008·SUPR1M3 opened Oct 28, 2026·4 tasks done·

  ·
  1

  1

• Preserve DeepSeek-V4 routing bias dtype across training steps

  #6007·janbernloehr opened Oct 28, 2026·4 tasks done·

  ·
  1