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
Assigned to nobody Loading
Sort

Pull requests list

Preserve configured activation in fused grouped experts ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4814 opened Sep 21, 2026 by taking-lying-flat Contributor Loading…
Use DistMuon with matrix-sharded compute for Kimi w13 CLA Signed This label is managed by the Meta Open Source bot.
#4811 opened Sep 21, 2026 by weifengpy Contributor Draft
Refactor model configuration and parallelization ownership ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4810 opened Sep 21, 2026 by tianyu-l Contributor Loading…
[DistMuon] Simplify layout planning and matrix views CLA Signed This label is managed by the Meta Open Source bot.
#4809 opened Sep 20, 2026 by weifengpy Contributor Draft
[config] Reject validation.steps=-1 when data-parallel degree > 1 ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4807 opened Sep 20, 2026 by YeonwooSung Contributor Loading…
2 tasks
[models] Skip torch.compile on the Muse Glimmer vision adapter ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4806 opened Sep 20, 2026 by YeonwooSung Contributor Loading…
2 tasks
[docs] Fix stale TrainingEngine, RL, and checkpoint config paths ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4805 opened Sep 20, 2026 by YeonwooSung Contributor Loading…
2 tasks
fix: do not pass positions to RoPE when attn_backend is sdpa ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl
#4804 opened Sep 20, 2026 by ysa2215 Loading…
[do not review] Support Qwen3.5 GDN CP ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4803 opened Sep 20, 2026 by acisseJZhong Contributor 4/4 Draft
Fix peak FLOPS lookup for non-H100 devices CLA Signed This label is managed by the Meta Open Source bot.
#4800 opened Sep 20, 2026 by HaoranYu111 Draft
[DistMuon] Support independently sharded variable-size blocks CLA Signed This label is managed by the Meta Open Source bot.
#4799 opened Sep 19, 2026 by weifengpy Contributor Draft
graph_trainer: two-level ILP memory policy CLA Signed This label is managed by the Meta Open Source bot.
#4793 opened Sep 18, 2026 by nnurlan008 Draft
[Kimi K3] Add packed MXFP4 import and MX QAT recipe ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4791 opened Sep 18, 2026 by elfiegg Contributor Draft
[RL] Restore model shard layouts after HF conversion ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4787 opened Sep 18, 2026 by drisspg Contributor 2/2 Loading…
[RL] Name the varlen attention backend after TorchTitan ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4786 opened Sep 18, 2026 by drisspg Contributor 1/2 Loading…
Fix DeepSeek V4 forward on pipeline stages ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4782 opened Sep 18, 2026 by HaoranYu111 Loading…
[DistMuon] Accelerate NS SYRK with optional QuACK symmetric GEMM ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4781 opened Sep 18, 2026 by WhoisZihan Loading…
[Kimi K3] Attention residual: zero initialisation and a recomputing aggregation ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4780 opened Sep 18, 2026 by QIU023 Contributor Loading…
Fix FSDP hanging for multimodal models ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4777 opened Sep 18, 2026 by jinsooihm Contributor Loading…
Add global batch-wise MoE auxiliary loss ciflow/fake-pg Run 1-GPU Fake PG integration tests ciflow/rl CLA Signed This label is managed by the Meta Open Source bot.
#4772 opened Sep 17, 2026 by pianpwk Contributor Loading…
[graph_trainer] WGrad fusion with grad accumulation ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4768 opened Sep 17, 2026 by IvanKobzarev Contributor Loading…
[DO NOT review, stack on K3 text PP PR #4312] [Kimi K3] PP ranks activation memory offload CLA Signed This label is managed by the Meta Open Source bot.
#4765 opened Sep 17, 2026 by QIU023 Contributor Draft
[DO NOT review, stack on K3 text PP PR #4312] [Kimi K3] cross PP ranks activation memory LoadBalancing CLA Signed This label is managed by the Meta Open Source bot.
#4764 opened Sep 17, 2026 by QIU023 Contributor Draft
[graph_trainer] Support runtime-aware graph passes ciflow/fake-pg Run 1-GPU Fake PG integration tests CLA Signed This label is managed by the Meta Open Source bot.
#4763 opened Sep 17, 2026 by qqq6op Loading…
Make DistMuon stream initialization device agnostic CLA Signed This label is managed by the Meta Open Source bot.
#4762 opened Sep 17, 2026 by weifengpy Contributor Draft
ProTip! Updated in the last three days: updated:>2026-09-17.