Repository Issue Activity (beta)

#4289 NoPE implementation and `max_seq_len`

#3029 [RFC] Kimi K3 Architecture support for torchtitan based pre-training and post-training (with veRL)

#4272 Kimi K3 issue tracking

#4293 SSRF: arbitrary URL fetch in image.py:36 without private IP validation

#2936 TODO Debt Report

#4273 MoE router gate's fp32 autocast forces a SIMT (non-tensor-core) GEMM: 4% of speed-of-light, ~8-9% of step time on B300

#4269 [Kimi K3] Add MXFP4-weight / MXFP8-activation QAT and checkpoint interoperability

#1118 Llama 4 issue tracking

#2544 [RFC] Support new models Qwen3.5-35B-A3B (text input only)

#4208 [RFC] Complete synchronous TorchFT integration

#4207 [RFC] Expose LocalSGD CPU staging in TorchTitan

#4209 [RFC] Make TorchFT checkpoint healing persistent and FSDP-safe

#4206 [RFC] Add an extensible TorchFT process-group registry

#4205 [Torchtitan][Pytorch][ROCm 7.14] Qwen3 MoE TP+EP training produces extremely large(nan) gradient norms

#4097 RL generator skips block_size=256 on Blackwell, crashing the FA2 fallback path

#4035 [RFC] DCP Checkpoint Load: Striped I/O + Collective Broadcast

#4149 spmd_types + AC + fake PG errors with wrong shapes

#4011 [RFC] Checkpoint Integrity Verification (SHA-256 Manifest)

#3353 Proposal: async and load-balanced Muon/SOAP-style optimizers for TorchTitan

#4125 CI misconfiguration prevents some jobs running on stack PRs

#3755 [rl] StickySessionRoutingStrategy can overload a generator for long multi-turn rollouts (no KV-budget awareness / no rebalancing)

#3622 [RL] MoE EP+TP: a vLLM-decode forward with fewer tokens than the TP degree breaks the sequence-parallel routing path

#933 [Feature] Add Multi-Token Prediction module

#3990 [feat] RFC: FP8 compilation support in GraphTrainer

#4010 [RFC] Support loose checkpoint loading via checkpoint.ignore_unexpected