All pull requests

Pull requests

• [GKD] Preserve EOS supervision for off-policy distillation

  #10285·taking-lying-flat opened Oct 2, 2026Contributor·1 of 4 tasks·

  ·
  1

• fix(template): handle assistant text before tool calls in ReAct templates and Seed-OSS

  #10284·Zhuoxi2000 opened Oct 1, 2026·1 of 4 tasks·

  ·

• fix: preserve Xing4.0 and MiniCPM5 tool argument types and whitespace

  #10281·taking-lying-flat opened Oct 1, 2026Contributor·1 of 4 tasks·

  ·
  2

• fix: respect Xing4.0 eager mode and reject conflicting patch options

  #10280·taking-lying-flat opened Oct 1, 2026Contributor·1 of 4 tasks·

  ·

• fix(ui): use the last assistant turn when marking a toolbench observation

  #10279·RerankerGuo opened Oct 1, 2026Contributor·

  ·

• fix(dataset): key the map cache on the arguments of the map

  #10278·RerankerGuo opened Oct 1, 2026Contributor·

  ·

• fix(dataset): restore dataset_infos.json after loading a local dataset folder

  #10277·RerankerGuo opened Oct 1, 2026Contributor·

  ·

• fix(megatron): defer full-vocabulary GKD teacher logits to forward_step

  #10276·FOWEPJF255 opened Oct 1, 2026·1 of 2 tasks·

  ·
  1

• fix(agent-template): add missing newline between assistant content and tool_call for glm4_5

  #10270·YueshenZ opened Sep 30, 2026Contributor·

  ·

• test(template): edge cases for mixed rollout response (#10077)

  #10269·RerankerGuo opened Sep 30, 2026Contributor·

  ·

• fix(agent): preserve Qwen tool argument types and whitespace

  #10267·taking-lying-flat opened Sep 30, 2026Contributor·1 of 4 tasks·

  ·

• fix(reward): follow the GRPO reward contract in MathORM

  #10266·Lesereingrape opened Sep 29, 2026Contributor·

  ·
  1

• fix(utils): report port-scan exhaustion instead of returning a busy port

  #10264·Lesereingrape opened Sep 29, 2026Contributor·1 of 4 tasks·

  ·
  1

• fix(infer): inline a local image whose path is longer than 200 characters

  #10262·SashaMIT opened Sep 29, 2026·

  ·

• fix(template): preserve original tool-call arguments JSON (#10216)

  #10261·RerankerGuo opened Sep 29, 2026Contributor·

  ·
  1

• fix(utils): restore ZeRO-3 hooks after generation errors

  #10260·fukalous opened Sep 29, 2026·1 of 4 tasks·

  ·
  1

• fix(template): preserve assistant reasoning on Qwen3.5/3.6 tool-call turns (#10255)

  #10259·RerankerGuo opened Sep 29, 2026Contributor·

  ·
  1

• perf(grpo): reuse rollout grouping for SP advantage updates

  #10257·taking-lying-flat opened Sep 29, 2026Contributor·1 of 4 tasks·

  ·

• fix(grpo): normalize Liger DAPO across gradient accumulation

  #10256·taking-lying-flat opened Sep 29, 2026Contributor·1 of 4 tasks·

  ·

• fix(agent_template): render tool call arguments as JSON in ReAct and GLM4 templates

  #10254·MohammadHijjawi97 opened Sep 28, 2026·1 of 4 tasks·

  ·

• fix(utils): treat a blank or non-numeric launch env var as unset

  discussion
  #10251·Lesereingrape opened Sep 27, 2026Contributor·1 of 4 tasks·

  ·

• [bugfix] drop dp_cp_group from content_metadata when saving mcore checkpoint

  discussionpending
  #10242·WhichPaths opened Sep 25, 2026·1 of 4 tasks·

  ·

• feat(megatron): add profiling support with pytorch/nsys profiler and memory history

  postpone
  #10237·HowardZorn opened Sep 23, 2026·2 of 4 tasks·

  ·
  5

• fix(infer): align TransformersEngine with doc defaults by propagating…

  invalid
  #10222·DreamBinary opened Sep 21, 2026·

  ·
  1

• Fix/tool user followup root

  postpone
  #10201·tastelikefeet opened Sep 19, 2026Collaborator·1 of 4 tasks·

  ·