All pull requests

Pull requests

• Avoid redundant source graphs in process tokenization results

  #1003·bradhilton opened Sep 27, 2026Collaborator·

  ·

• Price the adapter gradients a recompute backward holds at its peak

  #1002·bradhilton opened Sep 27, 2026Collaborator·

  ·

• Move TrainerRank checkpoint-slot and dynamic-optimizer management into trainer_rank/_slots and _optimizer

  #1001·bradhilton opened Sep 26, 2026Collaborator·

  ·
  2

• Move TrainerRank micro-batch planning into trainer_rank/_micro_batch_planner

  #1000·bradhilton opened Sep 26, 2026Collaborator·

  ·
  2

• Move TrainerRank memory accounting into trainer_rank/_memory

  #999·bradhilton opened Sep 26, 2026Collaborator·

  ·
  2

• Dispatch GPU image builds before fleet prewarm

  #995·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Avoid repeated numeric-list scans during trajectory serialization

  #990·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Fence lossy serial-tool history certificates by producer policy

  #989·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Reuse checkpoint snapshot headers and shard indexes per block

  #987·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Price dense recompute by its traced stage and one input gradient

  #986·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Reuse completed rendering during assistant boundary checks

  #985·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Read only needed checkpoint snapshot components during finalization

  #984·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Preserve forked native conditioning in multi-history tokenization

  #983·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Fix process-pool readiness under uneven worker startup

  #982·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Raise EP routed-row pricing to each checkpoint's observed routing

  #981·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Avoid per-node numeric type tuples in compact decoding

  #980·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Coalesce packed-expert optimizer checkpoint copies

  #979·bradhilton opened Sep 26, 2026Collaborator·

  ·

• Price CP2 recompute per rank on each layer type's layout

  #978·bradhilton opened Sep 25, 2026Collaborator·

  ·

• Reuse recorded tokens by default and preserve literal assistant content

  #977·bradhilton opened Sep 25, 2026Collaborator·

  ·

• Prove native conditioning before preserving literal reasoning markers

  #974·bradhilton opened Sep 25, 2026Collaborator·

  ·

• Try budgeted CUDA cache recovery before accepting internal splits

  #972·bradhilton opened Sep 25, 2026Collaborator·

  ·
  3

• Resume create_sft_dataset_iterator at the right chunk in later epochs

  #970·breken-ai opened Sep 25, 2026·

  ·

• Keep max_batch_exceptions for every batch in trajectory_group_batches

  #969·breken-ai opened Sep 25, 2026·

  ·

• Expose exact packing and input fingerprints in trainer telemetry

  #968·bradhilton opened Sep 25, 2026Collaborator·

  ·

• Reuse token alignment across chat history masks

  #965·bradhilton opened Sep 25, 2026Collaborator·

  ·