Issues · pytorch/torchtitan · GitHub

Labels Milestones

[graph_trainer] Graph-PP RFC

#3780

· sanketpurandare opened

on Jun 25, 2026

·

New Feature, Model, Misc

transformers_modeling_backend: pretrained dense models produce near-random loss / crash on current PyTorch

#3775

· HosseinKaviani-H opened

on Jun 24, 2026

8 GPU Model Tests red on main: gpt-oss PP+FSDP+EP+SACOP MoE _grouped_mm index OOB (CUDA scatter assert == ROCm HSA 0x1016)

#3761

· rishisinhanj opened

on Jun 24, 2026

[RL] Support Qwen3.5 in the RL experiment

#3756

· yichuan-w opened

on Jun 24, 2026

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

#3755

· yichuan-w opened

on Jun 24, 2026

vllm v2 model runner does not work with VLLMGenerator DP routing

#3752

· pzhan9 opened

on Jun 23, 2026

[rl] vLLM cudagraph: FULL corrupts generation, FULL_AND_PIECEWISE unsupported with the varlen attention backend

#3709

· yichuan-w opened

on Jun 18, 2026

[GraphTrainer] [AutoParallel] Tracker for outstanding issues

#3699

· aditvenk opened

on Jun 17, 2026

[rl] self.gate() aten dispatch to `bmm` instead of `mm`, causing extra batch-invariant `bmm`patching

#3670

· wwwjn opened

on Jun 15, 2026

[rl] vLLM torch.compile (VLLM_COMPILE) on the generator fails: inductor "out variant should have at least one out arg"

#3669

· yichuan-w opened

on Jun 15, 2026

Operator torchtitan.deterministic_topk.default does not have a sharding strategy registered.

#3645

· tianyu-l opened

on Jun 12, 2026

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

#3622

· wwwjn opened

on Jun 10, 2026