All pull requests

Pull requests

• fix(model): keep Qwen3-VL DeepStack gradients under activation recompute

  #48·celve opened Oct 19, 2026·

  ·
  1

• fix(peft): remove host syncs from the multi-LoRA forward and MoE slot routing

  #47·micahtyong opened Oct 12, 2026·

  ·

• feat(peft): support canonical MultiLoRA projections

  #46·yueming-yuan opened Oct 3, 2026Collaborator·

  ·

• Support shared-outer expert LoRA in MultiLoRA

  #45·yueming-yuan opened Oct 2, 2026Collaborator·

  ·

• tilelang MLA: cap the indexer top-k width at the sequence length

  #43·Arist12 opened Sep 28, 2026·

  ·

• fix(peft): support shared-expert and absorbed MLA Multi-LoRA

  #41·yueming-yuan opened Sep 22, 2026Collaborator·

  ·

• tilelang MLA: handle megatron-core's unfused DSA kv up-projection

  #40·Arist12 opened Sep 16, 2026·

  ·

• [model] fix: handle separate KV normalization in GLM TileLang attention

  bug
  #37·yushengsu-thu opened Sep 10, 2026Collaborator·

  ·

• [Do not merge - still verifiying]: GLM-5.3-Flash (glm5_next) bridge with KDA / DSA-kpool / mHC and a PEFT sequence-parallel fix

  #35·yushengsu-thu opened Sep 3, 2026Collaborator·

  ·
  10

• [model] fix: register both MTP submodule spellings in qwen3_next_bridge

  #11·Zhichenzzz opened Jun 10, 2026·

  ·

• add fix for qwen35 mtp layer name

  #2·artem-osmosis opened May 11, 2026·

  ·