-
Notifications
You must be signed in to change notification settings - Fork 983
Pull requests: pytorch/torchtitan
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
[ci] Run 8-GPU numerics when goldens change
ciflow/8gpu
CLA Signed
This label is managed by the Meta Open Source bot.
#4502
opened Sep 7, 2026 by
wwwjn
Contributor
Loading…
support CP for KimiK3
CLA Signed
This label is managed by the Meta Open Source bot.
#4500
opened Sep 6, 2026 by
acisseJZhong
Contributor
Loading…
[DO NOT review, stacked on K3 spmd declaration PR ] [Kimi K3] Tensor parallelism with sequence parallel support, on the spmd_types declarations
CLA Signed
This label is managed by the Meta Open Source bot.
Introduce torch_remat activation checkpointing
CLA Signed
This label is managed by the Meta Open Source bot.
#4497
opened Sep 6, 2026 by
anijain2305
Contributor
Loading…
[RFC] Support anticipatory routing to reduce MoE training loss spikes
CLA Signed
This label is managed by the Meta Open Source bot.
#4495
opened Sep 5, 2026 by
Dyslex7c
Loading…
[Kimi K3] Full spmd backend declaration for K3 CP, EP and TP/SP: the dense path, the tower over cp, the MoE seams, the multimodal inputs
CLA Signed
This label is managed by the Meta Open Source bot.
#4492
opened Sep 5, 2026 by
QIU023
Contributor
Loading…
[tests] Add DeepSeek V4 integration smoke
CLA Signed
This label is managed by the Meta Open Source bot.
#4490
opened Sep 5, 2026 by
YeonwooSung
Contributor
Loading…
1 of 3 tasks
Support MTP+PP
CLA Signed
This label is managed by the Meta Open Source bot.
#4488
opened Sep 5, 2026 by
acisseJZhong
Contributor
•
3/3
Loading…
preparing pipeline runtime for MTP+PP
CLA Signed
This label is managed by the Meta Open Source bot.
#4486
opened Sep 5, 2026 by
acisseJZhong
Contributor
•
2/3
Loading…
[nemotron_nano] Add Nemotron-3 hybrid Mamba-MoE model family support (4B, 31B, 120B, 550B)
CLA Signed
This label is managed by the Meta Open Source bot.
[gemma4] Add Gemma-4 model family support (12B & 31B)
CLA Signed
This label is managed by the Meta Open Source bot.
#4481
opened Sep 4, 2026 by
rpaulk
Loading…
GraphTrainer: expose activation-rematerialization region metadata
CLA Signed
This label is managed by the Meta Open Source bot.
#4479
opened Sep 4, 2026 by
wconstab
Contributor
Loading…
[graph_trainer] in-graph inplace gradient accumulation mode, to make it CudaGraph friendly
ciflow/h100.8
Trigger H100.8 CI
ciflow/rl
CLA Signed
This label is managed by the Meta Open Source bot.
#4477
opened Sep 4, 2026 by
IvanKobzarev
Contributor
Loading…
[graph_trainer] none AC MemoryPolicy, none Compilation mode
ciflow/h100.8
Trigger H100.8 CI
CLA Signed
This label is managed by the Meta Open Source bot.
#4476
opened Sep 4, 2026 by
IvanKobzarev
Contributor
Loading…
[CP] Add Ulysses CP for varlen attention
CLA Signed
This label is managed by the Meta Open Source bot.
#4459
opened Sep 4, 2026 by
fegin
Contributor
Loading…
[RL] Let varlen attention use cuDNN when eligible
ciflow/rl
CLA Signed
This label is managed by the Meta Open Source bot.
[metrics] Log host memory and CPU utilization
CLA Signed
This label is managed by the Meta Open Source bot.
#4457
opened Sep 4, 2026 by
ivy-zhou
Contributor
Loading…
Use even shard annotations for attention projections
CLA Signed
This label is managed by the Meta Open Source bot.
fused_mla: contiguous rope tiles, out-of-place Q rotation, deterministic launch config
CLA Signed
This label is managed by the Meta Open Source bot.
#4454
opened Sep 3, 2026 by
SherlockNoMad
Contributor
Loading…
[models] Add Qwen3.6 config registry
CLA Signed
This label is managed by the Meta Open Source bot.
#4453
opened Sep 3, 2026 by
wwwjn
Contributor
Loading…
Refactor Deepseek v4 to use Attention Gym's Compressed Sparse Attention
CLA Signed
This label is managed by the Meta Open Source bot.
#4452
opened Sep 3, 2026 by
floatingtrees
Contributor
•
Draft
support MTP with chunked loss
CLA Signed
This label is managed by the Meta Open Source bot.
#4451
opened Sep 3, 2026 by
acisseJZhong
Contributor
•
1/3
Loading…
[CP] Add Ulysses context parallel
CLA Signed
This label is managed by the Meta Open Source bot.
#4450
opened Sep 3, 2026 by
fegin
Contributor
Loading…
[CP] Gather K/V with spmd_types.redistribute
CLA Signed
This label is managed by the Meta Open Source bot.
#4449
opened Sep 3, 2026 by
fegin
Contributor
Loading…
[Kimi 2.7] Enable context parallel
CLA Signed
This label is managed by the Meta Open Source bot.
#4448
opened Sep 3, 2026 by
jinsooihm
Contributor
Loading…
Previous Next
ProTip!
Filter pull requests by the default branch with base:main.