All pull requests

Pull requests

• [workflow] improve the bug report issue template and add a regression template

  #6484·JazyJiang opened Oct 26, 2026Collaborator·

  ·
  1

• [booster] use the accelerator API for the TorchFSDP device id

  module: booster/plugin
  #6482·JazyJiang opened Oct 26, 2026Collaborator·

  ·
  1

• Restrict unpickler in async optimizer checkpoint loading (fixes #6471)

  #6478·08raviyadav opened Oct 22, 2026·1 of 11 tasks·

  ·

• [moe] add EPMoELayer with a dispatcher / experts data contract

  module: moe
  #6477·JazyJiang opened Oct 22, 2026Collaborator·

  ·

• [shardformer]: fix pooled token for left-padded inputs in pipeline sequence classification

  #6458·Arthur031221 opened Sep 29, 2026·7 of 11 tasks·

  ·
  2

• fix: use the accelerator API instead of hard-coded cuda/cpu device selection

  #6455·li-lizhe opened Sep 26, 2026·

  ·
  1

  1

• [shardformer]: fix llama causal mask with sequence parallelism and no flash attention

  #6452·LiRunGuo opened Sep 22, 2026·5 of 11 tasks·

  ·

• [shardformer]: fix qwen2/qwen3 causal mask with pipeline + sequence parallelism and no flash attention

  #6451·LiRunGuo opened Sep 22, 2026·5 of 11 tasks·

  ·
  1

• [shardformer]: support Qwen3-MoE with expert parallelism

  #6450·LiRunGuo opened Sep 22, 2026·8 of 11 tasks·

  ·

• [checkpoint]: unpad padded parameters when saving sharded MoE checkpoints

  #6449·LiRunGuo opened Sep 22, 2026·5 of 11 tasks·

  ·

• [tensor]: fix saving checkpoints of padded parameters with tp_size=1 (keep detach/clone of padded tensors unpaddable)

  #6448·LiRunGuo opened Sep 22, 2026·8 of 11 tasks·

  ·
  1

• [gemini]: fix loading checkpoints of models with tied parameters

  #6447·LiRunGuo opened Sep 21, 2026·8 of 11 tasks·

  ·
  1

• [shardformer]: fall back to native RMSNorm when apex is not installed

  #6446·LiRunGuo opened Sep 21, 2026·8 of 11 tasks·

  ·
  1

• [example]: fix Pytorch_DDP/Pytorch_ZeRO plans crashing in gpt gemini demo

  #6445·LiRunGuo opened Sep 21, 2026·3 of 11 tasks·

  ·

• [kernel]: fix Q/K cross-wiring and out-of-bounds accesses in fused_rotary_emb triton kernel

  #6444·LiRunGuo opened Sep 21, 2026·8 of 11 tasks·

  ·
  1

  1

• [fp8]: Read detected local world size environment variable

  #6438·tandede opened Aug 20, 2026·7 of 10 tasks·

  ·
  1

  1

• ci: add names to compatibility matrix steps for readability

  #6436·MGPOCKY opened Aug 1, 2026·2 tasks·

  ·

• Use HybridAdam for Gemini CPU optimizer offload

  #6435·sdjasj opened Jul 29, 2026·

  ·
  1

• Fix GPT-2 FlashAttention mask preparation without pipeline

  #6434·sdjasj opened Jul 29, 2026·

  ·
  1

• Preserve initialized optimizer state in low-level ZeRO

  #6433·sdjasj opened Jul 29, 2026·

  ·
  1

• Fix GaLore 8-bit updates with ZeRO CPU offload

  #6432·sdjasj opened Jul 29, 2026·

  ·
  1

• Fix LAMB norm reduction with ZeRO CPU offload

  #6431·sdjasj opened Jul 29, 2026·

  ·
  1

• docs: fix documentation typos

  #6422·Zhao73 opened Jun 1, 2026·

  ·

• fix infinite recursion in ColoTracer on newer PyTorch versions

  #6420·ssam18 opened Apr 19, 2026·

  ·
  1

• Fix: Handle div_scale when using Gemini plugin with CPUAdam

  #6415·Truong5724 opened Apr 15, 2026·10 of 11 tasks·

  ·
  1