All pull requests

Pull requests

• [fix][skycap] Token mode: match through client twins, so one full render costs one unbridged call

  #2461·xinyuangui2 opened Oct 10, 2026Collaborator·

  ·
  7

• [skyrl-gym][text2sql] Don't award reward for matching an empty gold result

  #2460·zhongerdeajie opened Oct 10, 2026·

  ·
  1

  4

• [tinker] Support tinker SDK 0.32/0.33 (pin <0.34)

  #2459·avigyabb opened Oct 10, 2026Collaborator·

  ·

• [Fix][Tinker] Make SkyRL-Train loss:sum compatible

  #2454·j316chuck opened Oct 9, 2026Contributor·

  ·
  3

• [skycap][bugfix] Give the parser the prompt, so a prefilled <think> stays out of content

  #2453·sunnyysetia opened Oct 9, 2026·

  ·
  4

• Notify all only on new capacity

  #2452·LazaroHurtado opened Oct 9, 2026·

  ·
  1

• [harbor_skycap] Images through Harbor, a sandbox concurrency cap and labels, SWE-bench Multimodal recipe

  #2450·xinyuangui2 opened Oct 9, 2026Collaborator·

  ·
  21

• [agent] add Linkup Rewind date-pinned search engine

  #2449·shauryajain21 opened Oct 8, 2026·

  ·

• reshuffle prompts each epoch in fully-async training

  #2448·coreyhu opened Oct 8, 2026·

  ·
  4

• save and restore each megatron rank's rng state in checkpoints

  #2447·coreyhu opened Oct 8, 2026·

  ·
  7

• name fully-async checkpoints after the last trained step

  #2446·coreyhu opened Oct 8, 2026·

  ·
  13

• apply temperature to fp32 logits in the torch fused lm-head path

  #2445·coreyhu opened Oct 8, 2026·

  ·
  1

• add summable cispo ratio metrics

  #2444·coreyhu opened Oct 8, 2026·

  ·

• write training batch dumps to cloud export paths

  #2443·coreyhu opened Oct 8, 2026·

  ·
  1

• download cloud directory contents into an existing local directory

  #2442·coreyhu opened Oct 8, 2026·

  ·
  4

• [skycap] skycap check: report whether a record stays on one path, and why each fork happened

  #2440·jadenfix opened Oct 7, 2026Contributor·2 tasks done·

  ·
  16

• [inference] Salt each LoRA adapter's prefix cache by load count instead of resetting every engine's cache on sync

  #2439·yapdianang opened Oct 7, 2026Contributor·

  ·
  8

• [megatron] KDA context parallelism: opt-in all-gather exchange (SKYRL_KDA_CP_EXCHANGE=allgather)

  #2437·avigyabb opened Oct 7, 2026Collaborator·2 tasks done·

  ·
  1

• [megatron][vlm] GLM-5.3-Flash vision-language model on Megatron (CP=1, packing)

  #2432·xinyuangui2 opened Oct 6, 2026Collaborator·

  ·

• [skycap] Images in token mode

  #2431·xinyuangui2 opened Oct 6, 2026Collaborator·

  ·
  15

• [fix][train] Per-field TensorList padding so padded rows stay self-consistent

  #2430·RaghuHemadri opened Oct 6, 2026·

  ·

• [megatron] GLM-5.3-Flash DSA: query-chunked fused sparse attention (opt-in)

  #2423·avigyabb opened Oct 5, 2026Collaborator·2 tasks done·

  ·
  6

• [sft] GLM-5.3-Flash long-context recipe and context-length sweep runner

  #2419·avigyabb opened Oct 5, 2026Collaborator·1 of 2 tasks·

  ·

• [megatron] Context parallelism for GLM-5.3-Flash (head-wise KDA, all-gather DSA)

  #2418·avigyabb opened Oct 5, 2026Collaborator·1 of 2 tasks·

  ·
  7

• [examples] GLM-5.3-Flash cluster setup for Anyscale: EFA and node-local uv caches

  #2417·avigyabb opened Oct 5, 2026Collaborator·2 tasks done·

  ·
  4