All pull requests

Pull requests

• [Feature][Training] Give the online DSpark draft its own learning rate

  #42·0z5a opened Oct 31, 2026Owner·

  ·

• [Bugfix] Fix runtime errors left by the 0z5a-branch sync merge

  #41·0z5a opened Oct 28, 2026Owner·

  ·

• [Feature][Training] Add an FSDP2 training backend

  #40·0z5a opened Oct 19, 2026Owner·

  ·

• [Rollout][Training] Speculative-decoding rollout benchmark, tail metrics and an MTP example

  #39·0z5a opened Oct 19, 2026Owner·

  ·

• [Bugfix][Training] Fix DSpark online draft training: grad-norm group and loss memory

  #38·0z5a opened Oct 19, 2026Owner·

  ·

• Prepare ordinary Ouro GRPO standard-backend acceptance

  #37·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: add initial systems paper and audited result tables

  #36·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: verify current H20 preparation and faster evidence transport

  #35·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: retain IQuest eager update gate failures

  #34·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: record three-family H20 CUDA update correctness

  #33·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: audit IQuest dual-attention replay

  #32·0z5a opened Oct 6, 2026Owner·

  ·

• ScaleRLT: verify streamed evidence archives

  #31·0z5a opened Oct 5, 2026Owner·

  ·

• ScaleRLT: preserve Huginn latent identity in prefix training

  #30·0z5a opened Oct 5, 2026Owner·

  ·

• ScaleRLT: extend prefix learner scheduling to Nanbeige

  #29·0z5a opened Oct 5, 2026Owner·

  ·

• ScaleRLT: checkpoint shared prefix and complete suffix waves

  #28·0z5a opened Oct 5, 2026Owner·

  ·

• Freeze matched B-stage native RLTT qualification profiles

  #27·0z5a opened Oct 5, 2026Owner·

  ·

• Bind differentiable prefix programs to the producing actor

  #26·0z5a opened Oct 5, 2026Owner·

  ·

• Add an opt-in Ouro prefix schedule to the MCore learner step

  #25·0z5a opened Oct 5, 2026Owner·

  ·

• Separate logical RLTT loss and plan actor prefix groups

  #24·0z5a opened Oct 5, 2026Owner·

  ·

• Bind native qualification to initial and published model tensors

  #23·0z5a opened Oct 5, 2026Owner·

  ·

• Bind native qualification to frozen source files

  #22·0z5a opened Oct 5, 2026Owner·

  ·

• Verify native rollout and held-out outputs after recovery

  #21·0z5a opened Oct 5, 2026Owner·

  ·

• Record native worker identities for resume qualification

  #20·0z5a opened Oct 5, 2026Owner·

  ·

• Audit native checkpoint recovery and useful Adam updates

  #19·0z5a opened Oct 5, 2026Owner·

  ·

• Prepare resident native reward qualification and signal audit

  #18·0z5a opened Oct 5, 2026Owner·

  ·