Pull requests: pytorch/torchtune

New pull request New

Pull requests list

Add support for distributed checkpointing of HF safetensors with DCP CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2851 openedJun 25, 2025by ankitageorge Loading…
1 of 8 tasks
 1
Enable activation offloading for XPU CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2847 openedJun 24, 2025by zxd1997066 Loading…
4 of 13 tasks
 1
[Gemma2] Use nn.SDPA via MultiHeadAttention CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2844 openedJun 20, 2025by Jack-Khuu • Draft
6 of 13 tasks
 1
[DO NOT MERGE] fix main CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2832 openedJun 16, 2025by joecummings Loading…
 1
[DONT MERGE] Debug branch for the Qwen3 + full model compile CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2831 openedJun 16, 2025by anijain2305 Loading…
 1
skip compiling opt step instead of erroring if opt_in_bwd=True CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2827 openedJun 13, 2025by felipemello1 Loading…
1 of 4 tasks
 3
raise error if fsdp_cpu_offload + opt_in_bwd CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2826 openedJun 13, 2025by felipemello1 Loading…
1 of 4 tasks
 1
Fix #2809, modify attention CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2822 openedJun 12, 2025by krammnic Loading…
2 of 13 tasks
1
2
[WIP] Qwen3 MoE support CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2820 openedJun 12, 2025by intervitens • Draft
6 tasks
 6
[RFC] on-the-fly packing CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2819 openedJun 12, 2025by felipemello1 Loading…
 14
[WIP] Integrate OptimizerInBackward into SFT distributed recipe CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2818 openedJun 11, 2025by joecummings • Draft
 1
Test alignment of shared methods in recipes CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2807 openedJun 9, 2025by Andrei-Aksionov • Draft
 1
Integrate Muon optimizer (2725) CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2803 openedJun 8, 2025by Saurabh750 Loading…
1 of 13 tasks
 13
Fix command in config CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2796 openedJun 6, 2025by krammnic Loading…
1 of 13 tasks
 7
[WIP] Proper tool calling support in the torchtune CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2794 openedJun 6, 2025by krammnic Loading…
2 of 13 tasks
 9
[RFC] Reward modeling CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2788 openedJun 5, 2025by krammnic Loading…
 1
[RFC] Iterable Dataset CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2785 openedJun 4, 2025by felipemello1 Loading…
 46
Ungate FP8 + TP CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2781 openedJun 3, 2025by nathan-az • Draft
3 of 13 tasks
 1
[NOT FOR REVIEW] Full knowledge distillation recipe TP + FP8 CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2767 openedMay 28, 2025by krammnic • Draft
13 tasks
 2
[WIP] DSV3 This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2764 openedMay 27, 2025by SalmanMohammadi • Draft
2 of 7 tasks
 2
Add LRScheduler.state_dict() to checkpoints CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2762 openedMay 23, 2025by omkar-334 Loading…
2 of 13 tasks
1
4
[WIP][DEBUG] llama4 debugging CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2756 openedMay 21, 2025by IvanKobzarev Loading…
 2
Fixing counting number of batches for accumulation through epoch CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2745 openedMay 17, 2025by wesbz Loading…
7 of 13 tasks
 6
Add feature ligerceloss CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2741 openedMay 16, 2025by mananchawla2005 Loading…
7 of 9 tasks
1
56
[WIP][Don't Review] Ray updates CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
#2724 openedMay 12, 2025by pbontrager • Draft
13 tasks
 1
ProTip! Follow long discussions with comments:>50.