Issues

• Auto-resume picks the model-only final checkpoint and fails with "Missing key ... dataloader.version"

  Bug

  pytorch/torchtitan#5054 · Filter by author David-Wu1119 opened on Oct 4, 2026
  

• [graph_trainer] Fuse Dist-MoE WGrad accumulation in functional graphs

  pytorch/torchtitan#5044 · Filter by author sanketpurandare opened on Oct 3, 2026
  1

• [graph_trainer] Preserve reduce-dtype accumulation for singleton FSDP gradients

  pytorch/torchtitan#5043 · Filter by author sanketpurandare opened on Oct 3, 2026
  1

• wandb silently drop validator log

  pytorch/torchtitan#5027 · Filter by author rakkit opened on Oct 2, 2026
  3

  

• NCCL all-reduce OOM despite reclaimable PyTorch cache during Kimi K3 training

  pytorch/torchtitan#5019 · Filter by author elfiegg opened on Oct 2, 2026
  

• [gpt_oss] HF export writes routed experts under MXFP4 *_blocks names, so transformers re-initializes every expert

  pytorch/torchtitan#4996 · Filter by author Zhuoxi2000 opened on Oct 1, 2026
  7

  

• GrainDataLoader DCP restore fails when Hugging Face streaming state changes structure

  pytorch/torchtitan#4991 · Filter by author da462 opened on Oct 1, 2026

• New torchtitan cannot fuse matmul reduce scatter when running AsyncTP

  pytorch/torchtitan#4937 · Filter by author jenniew opened on Sep 29, 2026
  5

  

• [RFC][rl] Restore the data stream and in-flight prompts when resuming RL training

  pytorch/torchtitan#4915 · Filter by author MarcelShi96 opened on Sep 27, 2026

• Context parallel does not cover Qwen3.5 Gated DeltaNet

  pytorch/torchtitan#4902 · Filter by author MegaGimen opened on Sep 26, 2026
  1

• TorchTitan model support for TorchComms backend

  wontfix
  pytorch/torchtitan#4839 · Filter by author newtdms opened on Sep 22, 2026
  1

• [CI] tests/unit_tests/rl is not collected by CPU/GPU unit jobs

  pytorch/torchtitan#4828 · Filter by author YeonwooSung opened on Sep 22, 2026
  1

• Restore the training loss for the DeepSeek V4 indexer

  pytorch/torchtitan#4802 · Filter by author HaoranYu111 opened on Sep 20, 2026
  2

• [transformers_modeling_backend]Regression of #3775: rotary inv_freq is left uninitialized after to_empty()

  pytorch/torchtitan#4783 · Filter by author Jiarwang77 opened on Sep 18, 2026
  3

• [Bug] An asynchronous TorchFT healing export reshards an in-flight FSDP2 parameter

  pytorch/torchtitan#4753 · Filter by author jojoinfra opened on Sep 16, 2026
  

• [RFC][DistMuon] Handle fused weights correctly for Kimi K3 per-head Muon

  pytorch/torchtitan#4692 · Filter by author JavaZeroo opened on Sep 15, 2026
  3

  

• [graph_trainer] SAR omits in-place writes to recomputed buffers, producing incorrect gradients

  pytorch/torchtitan#4688 · Filter by author johnjunjun7 opened on Sep 15, 2026
  1

• [MTP] Use per-module valid-token denominators for AuxLoss

  pytorch/torchtitan#4682 · Filter by author acisseJZhong opened on Sep 14, 2026
  1

• [graph_trainer][pipelining] Carry schedule metadata into reusable stage graphs

  pytorch/torchtitan#4655 · Filter by author sanketpurandare opened on Sep 14, 2026
  1

• [Bug][DistMuon] split fused gated linear weights before NS computation

  pytorch/torchtitan#4532 · Filter by author WhoisZihan opened on Sep 9, 2026

• [Bug] Weight initialized from the same seed when cpu offloading is active

  high prioritytriage review
  pytorch/torchtitan#4511 · Filter by author francesco-bertolotti opened on Sep 7, 2026
  1

  

• [Qwen3 MoE] Make auxiliary-loss-free expert-bias updates explicit and migration-safe

  pytorch/torchtitan#4394 · Filter by author yuweih205 opened on Sep 1, 2026
  5

• CUDAGraphable MoE Status

  pytorch/torchtitan#4367 · Filter by author syed-ahmed opened on Aug 28, 2026
  

• [RFC] Support Anticipatory Routing to mitigate MoE training loss spikes

  pytorch/torchtitan#4349 · Filter by author chango9543 opened on Aug 27, 2026
  6

• [Graph Trainer] [EP Overlap] Remove transformer-level chunking

  pytorch/torchtitan#4342 · Filter by author aditvenk opened on Aug 27, 2026