GitHub · Where software is built

Labels Milestones New issue

Expert group padding to multiple of 32

#2262

· eee4017 opened

on Jan 20, 2026

GPT-OSS MoE's implementation should use <code>route_norm=True</code>

#2258

· linyuhongg opened

on Jan 20, 2026

Weight tying not working

#2252

· francesco-bertolotti opened

on Jan 19, 2026

C4 Dataset Download Rate out of API limit

#2239

· zixianwang2022 opened

on Jan 15, 2026

Workflows being triggered on my fork

#2238

· aflah02 opened

on Jan 15, 2026

RuntimeError: Backend fake does not yet support sequence numbers.

#2236

· chennyso opened

on Jan 15, 2026

[Question] 2026 Roadmap and Development Scope for TorchTitan

#2233

· BIGBALLON opened

on Jan 15, 2026

Expert Parallel Fast Path and Router Dtype Restoration (from solar-open 102B)

#2225

· rakkit opened

on Jan 13, 2026

Efficiency Issue with Qwen3 attention module

#2224

· francesco-bertolotti opened

on Jan 13, 2026

Understanding RMSNorm Gradient Synchronization in Tensor-Parallel LLaMA

module: sequence parallel

module: tensor parallel

#2217

· alpemreacar opened

on Jan 9, 2026

Significant degradation of multi-stage Interleaved1F1B performance on MoE models

module: pipelining

#2209

· yuankaichen-amd opened

on Jan 8, 2026

Qwen3 shows different output when compiled vs not compiled

module: torch.compile

#2196

· halflearned opened

on Jan 5, 2026