Issues · pytorch/torchtitan · GitHub

Labels Milestones New issue

[CI] qwen3_vl_moe_fsdp+tp+ep OOM on g5.48xlarge (A10G 24GB)

module: flex attention

#3025

· acisseJZhong opened

on Apr 18, 2026

[graph_trainer] test_bitwise_deterministic hash expectations stale after nightly bump

#3024

· acisseJZhong opened

on Apr 18, 2026

graph_trainer JIT fails with 'Node tangents_2 was invalid, but is output' when module forward returns a DTensor

#3013

· fegin opened

on Apr 17, 2026

tyro CLI subcommand collision between Float8LinearConverter.Config and Float8GroupedMMConverter.Config breaks DeepSeek-V3-671B training

#2981

· janbernloehr opened

on Apr 15, 2026

PP and non-PP produce differently scaled gradients

#2972

· wwwjn opened

on Apr 15, 2026

Llama4 FP8 + EP: Dynamo recompile limit exceeded during SAC backward

module: torch.compile

#2958

· acisseJZhong opened

on Apr 14, 2026

[MoE] Token dispatcher and MoE cleanup TODOs

#2956

· acisseJZhong opened

on Apr 14, 2026

TODO Debt Report

#2936

· Andrei-Aksionov opened

on Apr 10, 2026

Document Fault Tolerant setup (TorchFT) timeouts

#2899

· yurigorokhov opened

on Apr 8, 2026

ChatDataset mid training resume not reproducible with multi epoch

#2864

· slimfrkha opened

on Apr 7, 2026

[graph_trainer] Nightly scout tracking issue

#2856

· SherlockNoMad opened

on Apr 7, 2026

"global_valid_tokens" scales badly

#2850

· rakkit opened

on Apr 4, 2026