Skip to content

Pull requests: pytorch/torchtitan

New pull request New
187 Open 1,505 Closed
187 Open 1,505 Closed
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
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

fix sdpa-varlen attention mismatch in qwen3 (#2229)
#2264 opened Jan 21, 2026 by tf170898 Loading…
2
[draft][LoRA] Add LoRA converter for LoRA finetuning ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2263 opened Jan 20, 2026 by mori360 • Draft
GQA without kv repeats CLA Signed This label is managed by the Meta Open Source bot.
#2259 opened Jan 20, 2026 by francesco-bertolotti Loading…
1
6
Remove unnecessary token padding for MoE in BF16 mode CLA Signed This label is managed by the Meta Open Source bot.
#2255 opened Jan 20, 2026 by rakkit Loading…
weight tying fix for qwen3 CLA Signed This label is managed by the Meta Open Source bot.
#2253 opened Jan 19, 2026 by francesco-bertolotti Loading…
1
7
[mxfp8 training] add new configurable params now exposed by torchao ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2251 opened Jan 18, 2026 by danielvegamyhre Loading…
4
[mxfp8 moe training] mxfp8 all to all ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2250 opened Jan 17, 2026 by danielvegamyhre Loading…
6
[mxfp8 moe training] support wgrad_with_hp recipe ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2249 opened Jan 17, 2026 by danielvegamyhre Loading…
2
Add ROCm CI support for Auto Parallel & Compiler Toolkit experiments ciflow/rocm-mi300 ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot. module: rocm
#2248 opened Jan 17, 2026 by akashveramd • Draft
@akashveramd
1
[rl] refactor grader and trainer generator actor ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2244 opened Jan 16, 2026 by wwwjn Loading…
8
[lint] ignore all existing pyrefly errors (v0.45.1) ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2241 opened Jan 16, 2026 by xmfan Loading…
[DONT LAND] Implement PrefetchedDataloader for overlapped data loading ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2232 opened Jan 14, 2026 by fegin Loading…
8
[MoE] Fix experts DTensor metadata bug for dcp ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2227 opened Jan 14, 2026 by shuhuayu • Draft
15
[rl] refactor save and load model weights using DCP ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2221 opened Jan 13, 2026 by wwwjn Loading…
5
Added ROCm CI support for simple fsdp & torchcomms experiments test ciflow/rocm-mi300 ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot. module: rocm
#2220 opened Jan 12, 2026 by akashveramd • Draft
@akashveramd
1
feat(gpt-oss): add YaRN RoPE extensions with mscale for extended context CLA Signed This label is managed by the Meta Open Source bot.
#2216 opened Jan 8, 2026 by eous Loading…
4
feat(training): add freeze_router_bias and freeze_expert_bias configs… CLA Signed This label is managed by the Meta Open Source bot.
#2215 opened Jan 8, 2026 by eous Loading…
3
fix: enable torch.autocast for TP parallelism without FSDP CLA Signed This label is managed by the Meta Open Source bot.
#2213 opened Jan 8, 2026 by eous Loading…
3
feat(moe): add topk_before_score routing and use_router_bias support CLA Signed This label is managed by the Meta Open Source bot.
#2212 opened Jan 8, 2026 by eous Loading…
@shuhuayu
6
fix(gpt-oss): correct attention sink from sigmoid to LSE renormalization CLA Signed This label is managed by the Meta Open Source bot.
#2211 opened Jan 8, 2026 by eous Loading…
@wwwjn
4
feat: add differential learning rate and weight decay support CLA Signed This label is managed by the Meta Open Source bot.
#2210 opened Jan 8, 2026 by eous Loading…
11
[HybridEP] Support hybridEP for GB200 with NVL72 CLA Signed This label is managed by the Meta Open Source bot.
#2207 opened Jan 8, 2026 by elfiegg Loading…
5
Fix loss computation by handling valid token imbalance in train loop ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot.
#2206 opened Jan 7, 2026 by wwwjn Loading…
62
feat(gpt-oss): Add CPU offload optimizer, differential LR/WD, and more CLA Signed This label is managed by the Meta Open Source bot.
#2205 opened Jan 7, 2026 by eous Loading…
31
Add ROCm support for H100 tests ciflow/rocm-mi300 ciflow/8gpu CLA Signed This label is managed by the Meta Open Source bot. module: rocm
#2202 opened Jan 5, 2026 by akashveramd • Draft
@akashveramd
4
ProTip! Exclude everything labeled bug with -label:bug.