Skip to content

Pull requests: pytorch/torchtitan

Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Assignee
Filter by who’s assigned
Sort

Pull requests list

update performance and loss converging results CLA Signed This label is managed by the Meta Open Source bot.
#800 opened Jan 22, 2025 by tianyu-l Loading…
Only init the gloo process group when necessary CLA Signed This label is managed by the Meta Open Source bot.
#798 opened Jan 20, 2025 by carmocca Loading…
[cp] Add cudnn attention support to Context Parallel CLA Signed This label is managed by the Meta Open Source bot.
#796 opened Jan 17, 2025 by XilunWu Draft
[debug only] microbatching CLA Signed This label is managed by the Meta Open Source bot.
#795 opened Jan 16, 2025 by tianyu-l Draft
[BE] Lr schduler flatten CLA Signed This label is managed by the Meta Open Source bot.
#794 opened Jan 16, 2025 by mori360 Loading…
[do NOT land] CP+torch.compile debugging attempt CLA Signed This label is managed by the Meta Open Source bot.
#791 opened Jan 15, 2025 by XilunWu Draft
Make CheckpointManager friendlier to custom StorageWriter/StorageReader CLA Signed This label is managed by the Meta Open Source bot.
#789 opened Jan 12, 2025 by dimdi-y Loading…
[WIP] support zbv CLA Signed This label is managed by the Meta Open Source bot.
#787 opened Jan 10, 2025 by H-Huang Draft
Register backward hook for the whole optim_dict to enable working at multi schedule pp CLA Signed This label is managed by the Meta Open Source bot.
#780 opened Jan 7, 2025 by mori360 Draft
[Not for land] Integrate float8nocompile, an experimental feature for high performance CLA Signed This label is managed by the Meta Open Source bot.
#778 opened Jan 7, 2025 by danielvegamyhre Loading…
[PoC] Typed JobConfig CLA Signed This label is managed by the Meta Open Source bot.
#767 opened Jan 1, 2025 by jaysonfrancis Loading…
[MoE][PoC] Expert Parallel: dp2ep CLA Signed This label is managed by the Meta Open Source bot.
#732 opened Dec 12, 2024 by tianyu-l Draft
[MoE][PoC] Expert Parallel: tp and tp2ep CLA Signed This label is managed by the Meta Open Source bot.
#731 opened Dec 12, 2024 by tianyu-l Draft
[MoE][PoC] model code CLA Signed This label is managed by the Meta Open Source bot.
#730 opened Dec 12, 2024 by tianyu-l Draft
[Not for land] Show replicated fp32 norm weights CLA Signed This label is managed by the Meta Open Source bot.
#717 opened Dec 4, 2024 by awgu Draft
First draft Auto-SAC workflow CLA Signed This label is managed by the Meta Open Source bot.
#710 opened Dec 2, 2024 by sanketpurandare Draft
[WIP] Allow benchmark between multiple configs CLA Signed This label is managed by the Meta Open Source bot.
#703 opened Nov 26, 2024 by H-Huang Loading…
[WIP] Adding OBELICS DataLoader CLA Signed This label is managed by the Meta Open Source bot.
#663 opened Oct 30, 2024 by TJ-Solergibert Loading…
[not for land] torch.compile individual linears CLA Signed This label is managed by the Meta Open Source bot.
#661 opened Oct 29, 2024 by vkuzo Loading…
Use enable_gqa in place of repeat_kv CLA Signed This label is managed by the Meta Open Source bot.
#641 opened Oct 22, 2024 by awgu Draft
Init weights only if not loading a checkpoint CLA Signed This label is managed by the Meta Open Source bot.
#628 opened Oct 18, 2024 by carmocca Draft
[DO NOT REVIEW] gaps to enable FDSP2 cpu offloading CLA Signed This label is managed by the Meta Open Source bot.
#622 opened Oct 16, 2024 by weifengpy Loading…
[Not for land] Settings to make Llama3-8B on 8 GPUs faster CLA Signed This label is managed by the Meta Open Source bot.
#615 opened Oct 14, 2024 by awgu Draft
[not for land] TE experiments, take 2 CLA Signed This label is managed by the Meta Open Source bot.
#614 opened Oct 14, 2024 by vkuzo Loading…
[DO NOT REVIEW] --experimental.fsdp_sharding_on_largest_dim CLA Signed This label is managed by the Meta Open Source bot.
#607 opened Oct 9, 2024 by weifengpy Loading…
ProTip! no:milestone will show everything without a milestone.