All pull requests

Pull requests

• [Feature] Support versioned Mooncake Store caching for LMDeploy RL

  #2290·RunningLeon opened Nov 10, 2026Contributor·8 of 10 tasks·

  ·

• [Feature] Preserve external frozen MTP across RL updates and checkpoints

  #2289·xs1997zju opened Nov 10, 2026Collaborator·

  ·

• [Feat] Add Dense MTP and Qwen3.5 VL Dense 4B/27B support on current main

  #2285·matrix72c opened Nov 9, 2026Contributor·5 tasks·

  ·

• [Test] Extend sequence-parallel logprob numerical coverage

  #2283·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Avoid waiting for failed training batch cleanup

  #2282·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Release unused device cache between training roles

  #2281·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Include FSDP strided shards in gradient norms

  #2280·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Accept assistant templates without an EOS suffix

  #2279·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Release recursive HF loader closure after loading

  #2278·matrix72c opened Nov 9, 2026Contributor·

  ·

• [Fix] Pool pinned model and optimizer storage with checkpoint-safe DTensor views

  #2277·matrix72c opened Nov 9, 2026Contributor·

  ·

• Slim the training-bound rollout payload at the object-store boundary

  #2276·ejunered opened Nov 9, 2026Collaborator·

  ·

• Overlap train-step CPU preparation with rollout offload and trajector…

  #2275·ejunered opened Nov 9, 2026Collaborator·

  ·

• [Fix] Reschedule LR when a resumed run changes its length

  #2272·jayhenry opened Nov 6, 2026Collaborator·

  ·

• [Test] Refresh fixtures broken by the sandbox user and kpool selector changes

  #2271·jayhenry opened Nov 6, 2026Collaborator·

  ·

• [Feature] Add opt-in FP32 QKV intermediates for KDA

  #2270·RunningLeon opened Nov 6, 2026Contributor·

  ·

• [Feature] Reuse LMDeploy KPool kernels in GLM-5.3 flash indexer

  #2269·qescccczmr opened Nov 6, 2026Collaborator·

  ·
  4

• [ci] remove some cases

  #2268·kkscilife opened Nov 6, 2026Collaborator·

  ·

• [ ci ]add glm5.3-flash rl case

  #2267·kkscilife opened Nov 6, 2026Collaborator·

  ·

• [Feature] Add critic value-clip and session value-health metrics

  #2264·matrix72c opened Nov 6, 2026Contributor·

  ·
  1

• [Fix] Use Ray reference counting for routed-expert lifetime

  #2263·matrix72c opened Nov 6, 2026Contributor·

  ·
  1

• [Fix] Onload the role model before the HF shard save

  #2262·matrix72c opened Nov 6, 2026Contributor·

  ·
  1

• [Fix] Preserve FP32 coefficients and accumulation in mHC post

  #2260·RunningLeon opened Nov 5, 2026Contributor·

  ·

• [Fix] Prevent hangs when tokenizer workers fail

  #2259·bingdongni opened Nov 5, 2026·

  ·

• [Draft] Retain native Agent tasks across complete-turn weight publications

  #2257·xs1997zju opened Nov 5, 2026Collaborator·

  ·

• [Fix] Preserve truncated Anthropic tool arguments in traces

  #2255·matrix72c opened Nov 5, 2026Contributor·

  ·
  1