All pull requests

Pull requests

• Add native score-centered MIS and asynchronous mixed-sampler RL

  #133·kailash109 opened Oct 25, 2026Collaborator·

  ·

• Bump rollout sglang to v0.5.22 and digest-pin

  #132·andrewhinh opened Oct 25, 2026Collaborator·

  ·
  1

  

• Digest-pin miles megatron base image

  #131·andrewhinh opened Oct 25, 2026Collaborator·

  ·
  1

  

• [DS-v4.1-Flash support] Add DeepSeek-V4.1-Flash-1017 LoRA recipes

  #129·micahtyong opened Oct 24, 2026Collaborator·

  ·

• [DS-v4.1-Flash support] One weight download, B200 and H200 fallbacks, and a 32K recipe

  #128·micahtyong opened Oct 24, 2026Collaborator·

  ·

• [security] Require authenticated Modal Servers for HTTP endpoints

  #127·jvmncs opened Oct 24, 2026Contributor·

  ·

• Add experimental TorchTitan multi-LoRA backend

  #126·kailash109 opened Oct 23, 2026Collaborator·

  ·

• [routing 4/4] Integrate native routing with exact-version sampling

  #123·kailash109 opened Oct 23, 2026Collaborator·

  ·

• [routing 3/4] Select LoRA workers by KV reuse and memory pressure

  #122·kailash109 opened Oct 23, 2026Collaborator·

  ·

• [routing 2/4] Publish incremental cache observations off the scheduler

  #121·kailash109 opened Oct 23, 2026Collaborator·

  ·

• [routing 1/4] Replay duplicate generations and bound cancellation

  #120·kailash109 opened Oct 23, 2026Collaborator·

  ·
  3

• Document the Miles LoRA backend and supported training backends

  #117·kailash109 opened Oct 22, 2026Collaborator·

  ·

• Add user app tags to recipes and scoped runs for cost attribution

  #111·micahtyong opened Oct 21, 2026Collaborator·

  ·

• Run every Miles recipe on Miles v0.1.2

  #110·micahtyong opened Oct 21, 2026Collaborator·

  ·

• Fix sampler warm-state routing in Modal deployments

  #108·zhouhelena1 opened Oct 21, 2026·

  ·

• Enable output-head LoRA for five additional model recipes

  #106·kailash109 opened Oct 20, 2026Collaborator·

  ·

• [cost attribution] Per-run trainer and sampler apps tagged with run metadata

  #105·shariqm-modal opened Oct 20, 2026·

  ·
  1

• Add opt-in score-centering loss

  #100·kailash109 opened Oct 19, 2026Collaborator·

  ·

• [cold start] Download weights while the sampler pool deploys

  #96·micahtyong opened Oct 18, 2026Collaborator·

  ·

• Support the Dimuon optimizer on the Miles LoRA backend

  #85·jvmncs opened Oct 16, 2026Contributor·

  ·

• Preserve original 27B DAPO experiment and runtime changes

  #84·kailash109 opened Oct 16, 2026Collaborator·

  ·

• Add extensible LoRA routing with model placement and overflow

  #72·kailash109 opened Oct 16, 2026Collaborator·

  ·

• Design: FP8 training, weight export, and inference

  #71·kailash109 opened Oct 15, 2026Collaborator·

  ·

• Add the DeepSeek-V4.1-Flash full-parameter recipe

  #66·micahtyong opened Oct 15, 2026Collaborator·4/4·

  ·

• Support large full-parameter models: local disk, BF16 asset conversion, staged sampler updates

  #65·micahtyong opened Oct 15, 2026Collaborator·3/4·

  ·