GitHub · Where software is built

Labels Milestones New issue

Potentially incorrect attention flop calculation due to wrong head_dim?

#1920

· gau-nernst opened

on Oct 18, 2025

Benchmark SymmMem's all_to_all_vdev_2d on NVL72

#1914

· syed-ahmed opened

on Oct 17, 2025

checkpoint.initial_load_in_hf should overwrite everything and load from hf weights.

#1900

· wukaixingxp opened

on Oct 16, 2025

[qwen3] Loss is extremely hight when initialized with seed checkpoint

#1879

· wwwjn opened

on Oct 15, 2025

After applying commit 6a3a9da, the training job runs out of GPU memory (OOM).

#1855

· limou102 opened

on Oct 11, 2025

Pipeline Parallelism Across Nodes Fails with EOFError

#1852

· rrutmann opened

on Oct 10, 2025

More robust distributed loss

#1842

· lkhphuc opened

on Oct 9, 2025

Deepseekv3 CUDA illegal memory access on B200 from increasing seq_len to 8192

#1791

· danielvegamyhre opened

on Oct 3, 2025

How to add supervised finetuning mask in torchtitan?

#1781

· kailashg26 opened

on Oct 1, 2025

Posttraining Library

#1771

· MarkLiLabs opened

on Sep 30, 2025

unwaited collective in DSV3+FSDP2+TP+compile

module: torch.compile

#1769

· ruisizhang123 opened

on Sep 29, 2025

Async TP CI failing

ci test failure

#1757

· fegin opened

on Sep 25, 2025