All pull requests

Pull requests

• [fix] Demote HTTPX request logs in Ray workers

  #2480·SumanthRH opened Oct 14, 2026Member·

  ·
  5

• [chore] Use random tokens in the full context trainer's dummy batch

  #2479·SumanthRH opened Oct 14, 2026Member·

  ·

• [harbor] Run Harbor's modal environment on Ray Sandbox

  #2478·xyuzh opened Oct 13, 2026·

  ·
  10

• [feat][megatron] Add an opt-in streaming write for async checkpoints

  #2477·bvolpato opened Oct 13, 2026Contributor·

  ·

• [feat][train] Honor return_per_token_outputs in the RL loss

  #2476·bvolpato opened Oct 13, 2026Contributor·

  ·

• [feat][train] Run fully async checkpoint and HF export during eval

  #2475·bvolpato opened Oct 13, 2026Contributor·

  ·

• fix(examples): align custom policy loss with trainer contract

  #2474·raghulchandramouli opened Oct 13, 2026·

  ·
  2

• [r3] Routed-experts capture with vLLM pipeline parallelism (runtime relay patch)

  #2472·SumanthRH opened Oct 13, 2026Member·2/2·

  ·
  7

• [fix] Enable pipeline-parallel router replay

  #2471·jeffreywang88 opened Oct 13, 2026·

  ·
  5

  

• [chore] Bump vLLM to 0.31.0, flashinfer 0.7.0.post1, Megatron-Bridge/Core, cutlass-dsl, xgrammar

  run_gpu_cirun_h100_gpu_cirun_megatron_gpu_ci
  #2470·SumanthRH opened Oct 13, 2026Member·1/2·

  ·
  3

• [vllm][lora] Size MoE expert LoRA slots at the expert adapter rank with normalize_moe_lora=true

  #2469·SumanthRH opened Oct 13, 2026Member·

  ·

• [megatron] mHC layer: honor mlp_chunks_for_training (chunk the MLP/MoE over the sequence)

  #2468·avigyabb opened Oct 12, 2026Collaborator·2 tasks done·

  ·
  4

• [harbor] Ray Sandbox environment backend (experimental, Ray nightly)

  #2467·xinyuangui2 opened Oct 12, 2026Collaborator·

  ·

• [tinker] Reproduction harness for the train_attn/train_mlp results (not for merge)

  #2465·ArturNiederfahrenhorst opened Oct 12, 2026·

  ·

• [tinker] Honour train_attn/train_mlp from Tinker's LoraConfig on the skyrl-train backend

  #2464·ArturNiederfahrenhorst opened Oct 12, 2026·

  ·
  1

  24

• [feat][fully-async] Support TrainingCallback in FullyAsyncRayPPOTrainer

  #2463·kar137 opened Oct 10, 2026·

  ·
  7

• [skycap] Token mode: bridge a turn the model ended without closing its thinking

  #2462·xinyuangui2 opened Oct 10, 2026Collaborator·

  ·

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

  #2461·xinyuangui2 opened Oct 10, 2026Collaborator·

  ·
  11

  

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

  #2460·zhongerdeajie opened Oct 10, 2026·

  ·
  1

  9

• [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·

  ·
  47

  

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

  #2449·shauryajain21 opened Oct 8, 2026·

  ·