Issues

• [RFC] Let custom routed-expert layouts declare their FSDP shard dimension

  pytorch/torchtitan#5545 · Filter by author samansg-aws opened on Oct 19, 2026
  4

• [qwen3] No local_compile region covers the block's norm/residual seams

  pytorch/torchtitan#5537 · Filter by author yuxinwan-amd opened on Oct 19, 2026
  3

• get_cuda_flash_attention_impl() selects FA3/FA4 by compute capability alone, so attention init raises ModuleNotFoundError when the wheel is absent

  pytorch/torchtitan#5536 · Filter by author janbernloehr opened on Oct 19, 2026

• [RFC] Device-agnostic training flows and tests

  pytorch/torchtitan#5502 · Filter by author JavaZeroo opened on Oct 18, 2026

• [RFC][rl] GRPO on Intel XPU: torchtitan + Monarch + TorchStore

  pytorch/torchtitan#5401 · Filter by author songhappy opened on Oct 15, 2026
  3

• [CI] GraphTrainer EP recipes retain pre-EP sharding after config construction change (#5280)

  pytorch/torchtitan#5361 · Filter by author sanketpurandare opened on Oct 13, 2026
  2

  

• [CI] GPT-OSS and Qwen3 model goldens fail after batch-wise MoE balancing (#4772)

  pytorch/torchtitan#5360 · Filter by author sanketpurandare opened on Oct 13, 2026
  1

  

• [DistMuon] Performance and memory optimization

  pytorch/torchtitan#5278 · Filter by author JavaZeroo opened on Oct 11, 2026

• [bug] gpt-oss quantized weight loading broken

  pytorch/torchtitan#5189 · Filter by author haydn-jones opened on Oct 9, 2026
  3

  

• [remat] SelectiveAC replay exposes attention projection placeholders to RoPE

  pytorch/torchtitan#5173 · Filter by author sanketpurandare opened on Oct 8, 2026
  

• [RFC] Integrate NCCL RMA into TorchTitan for zero-SM pipeline parallel communication

  pytorch/torchtitan#5150 · Filter by author DAMI211 opened on Oct 8, 2026
  

• CUDAGraph + Copy-on-write tensor lazy copy produce NaN output

  pytorch/torchtitan#5125 · Filter by author wwwjn opened on Oct 7, 2026
  

• DCP explicit load_step accepts incomplete checkpoint dirs and fails opaquely

  pytorch/torchtitan#5096 · Filter by author chinmayirm opened on Oct 6, 2026
  3

  

• 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
  2

  

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

  pytorch/torchtitan#5043 · Filter by author sanketpurandare opened on Oct 3, 2026
  2

  

• 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
  15

  

• [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

  

• New torchtitan cannot fuse matmul reduce scatter when running AsyncTP

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

  

• 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

• [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