GitHub · Where software is built

Labels Milestones New issue

[CI] Investigate LLVM failure

#2722

· Lucaskabela opened

on Mar 26, 2026

Change default attn backend to varlen/flex attention

#2720

· joecummings opened

on Mar 26, 2026

[RL] Support flex attention kernel for generator and trainer.

#2719

· zhxchen17 opened

on Mar 26, 2026

[RL] Unify mask creation logic for RL.

#2716

· zhxchen17 opened

on Mar 26, 2026

[RL] Dynamic Batching in RL loop

#2693

· wwwjn opened

on Mar 24, 2026

[RL] Loss function refactor in RL loop

#2692

· wwwjn opened

on Mar 24, 2026

[RL] Add CI to guard on numerics change

#2691

· wwwjn opened

on Mar 24, 2026

Scale up RL to Qwen3 32B, Qwen3 30B (MoE), and Deepseek V3

#2681

· daniellepintz opened

on Mar 24, 2026

[RL][Bug] Move towards batched compute_log_probs to improve efficiency

#2674

· Lucaskabela opened

on Mar 23, 2026

[RL] Torch.compile() gives logit difference post compilation as opposed to `.compile()`

#2673

· Lucaskabela opened

on Mar 23, 2026

Default validator skips a validation batch each round

#2636

· albertoveneri opened

on Mar 20, 2026

DCP loading fails when `torch.compile` target is nested inside `CheckpointWrapper`

#2622

· man2machine opened

on Mar 18, 2026