All pull requests

Pull requests

• Preserve native dtype in distributed group_mean instead of casting to float32

  #1103·OnePunchMonk opened Oct 3, 2026·1 of 2 tasks·

  ·
  1

• perf: reuse MoE terms within one memory evaluation

  #1101·bradhilton opened Oct 3, 2026Collaborator·

  ·

• fix(trainer-rank): credit reusable cache when placing gathered outputs

  #1100·bradhilton opened Oct 3, 2026Collaborator·

  ·

• fix: sum nonfused column LoRA input gradients

  #1092·bradhilton opened Oct 3, 2026Collaborator·

  ·
  1

• fix(trainer-rank): price dense CP2 recompute and no-grad waves from traced stages

  #1090·bradhilton opened Oct 3, 2026Collaborator·

  ·

• fix: preserve input gradients in fused LoRA layers

  #1087·bradhilton opened Oct 3, 2026Collaborator·

  ·

• perf(megatron): reuse packed weights for frozen MoE experts

  #1061·bradhilton opened Oct 2, 2026Collaborator·

  ·

• perf(trainer-rank): try budgeted cache recovery before splitting

  #1053·bradhilton opened Sep 30, 2026Collaborator·

  ·
  2

• fix: preserve container state and dictionary order during compaction

  #1015·bradhilton opened Sep 28, 2026Collaborator·

  ·

• fix(ruler): add score normalization, clamping, and robust JSON extraction (#477)

  #1004·shobhitagnihotri69 opened Sep 28, 2026·

  ·
  1

• perf: avoid repeated numeric-list scans during serialization

  #990·bradhilton opened Sep 26, 2026Collaborator·

  ·

• fix(trainer-rank): price dense CP2 recompute from traced stages

  #986·bradhilton opened Sep 26, 2026Collaborator·

  ·

• perf: reuse completed rendering during assistant boundary checks

  #985·bradhilton opened Sep 26, 2026Collaborator·

  ·

• fix: preserve forked native conditioning in multi-history tokenization

  #983·bradhilton opened Sep 26, 2026Collaborator·

  ·

• fix: verify distinct process-pool workers during startup

  #982·bradhilton opened Sep 26, 2026Collaborator·

  ·

• fix(trainer-rank): raise routed-row estimates from checkpoint observations

  #981·bradhilton opened Sep 26, 2026Collaborator·

  ·

• fix(trainer-rank): price CP2 recompute using each rank’s layouts

  #978·bradhilton opened Sep 25, 2026Collaborator·

  ·

• fix: reuse recorded tokens by default and preserve assistant content

  #977·bradhilton opened Sep 25, 2026Collaborator·

  ·

• fix: prove native conditioning before preserving reasoning markers

  #974·bradhilton opened Sep 25, 2026Collaborator·

  ·

• 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·

  ·

• fix(trainer-rank): account for recomputed mixer memory and head stages

  #963·bradhilton opened Sep 25, 2026Collaborator·

  ·

• fix: preserve sampled conditioning during exact history recovery

  #957·bradhilton opened Sep 25, 2026Collaborator·

  ·
  1

• perf: reuse cumulative counts for planner length thresholds

  #956·bradhilton opened Sep 25, 2026Collaborator·

  ·

• fix: honor oversized admission for supported unsplit EP plans

  #953·bradhilton opened Sep 25, 2026Collaborator·

  ·