Issues · pytorch/torchtitan · GitHub

Labels Milestones New issue

FlexAttention `BlockMask` breaks Pipeline Parallelism with split-backward (zero-bubble / multi) schedules

module: flex attention

module: pipelining

#3584

· tianyu-l opened

on Jun 9, 2026

[RL] vLLM attention masks life cycle

module: flex attention

#3572

· tianyu-l opened

on Jun 8, 2026

checkpoint interoperability issue between fused and non-fused Module definition

#3569

· tianyu-l opened

on Jun 7, 2026

[RL] titan GRPO batch invariant logbrob diff != 0

#3498

· liangel-02 opened

on Jun 3, 2026

[RL][Flex] batch invariant Flex Attention doesn't work with general compile

#3497

· liangel-02 opened

on Jun 3, 2026

[MoE] Qwen3 MoE with TP+EP fails: "histogram_cpu" not implemented for torch.int64

#3465

· yichuan-w opened

on May 31, 2026

HybridEP DispatchHandle not working with torch.compile

module: torch.compile

#3439

· chelsea0x3b opened

on May 28, 2026

`LoRAConverter` silently leaves non-Linear parameters trainable

#3438

· yh8899 opened

on May 28, 2026

[gpt_oss] Avoid materialising `(x_linear + 1)` intermediate in `swiglu` to reduce activation memory

#3437

· janbernloehr opened

on May 27, 2026

RFC: Graph-Based EP Overlap For GraphTrainer

#3434

· sanketpurandare opened

on May 27, 2026

[TensorParallel] NoParallel does not handle multi-output modules

#3374

· aditvenk opened

on May 16, 2026

[question] Why is the load balancing disabled for multi-mask models

#3368

· francesco-bertolotti opened

on May 15, 2026