12 Pull requests merged by 2 people

• consolidate model parallelize code between train modules

  #196 merged Mar 8, 2025

• make initialization more robust for custom modules

  #195 merged Mar 7, 2025

• FSDP optimizations

  #194 merged Mar 7, 2025

• Some fixes

  #193 merged Mar 7, 2025

• MoE parallelism improvements

  #192 merged Mar 6, 2025

• Steps per epoch convenience

  #190 merged Mar 6, 2025

• Add support for BF16 optim state in SkipStepAdamW

  #148 merged Mar 5, 2025

• Add quick settings to disable checkpoints and evals

  #189 merged Mar 4, 2025

• remove use of deprecated tensor.storage()

  #188 merged Mar 4, 2025

• update minimum torch version to 2.6.0

  #187 merged Mar 4, 2025

• Improve TP/CP APIs, add a fused linear loss

  #185 merged Mar 4, 2025

• Add support for context parallelism (round 2)

  #175 merged Mar 2, 2025

2 Pull requests opened by 2 people

• Add option to save data paths

  #191 opened Mar 6, 2025

• Add saving and loading of HF checkpoints

  #197 opened Mar 8, 2025

1 Issue closed by 1 person

• `get_total_norm` requires `torch==2.6.0`

  #186 closed Mar 4, 2025

2 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• making converter additionally support deepseek-coder dense hf model

  #167 commented on Mar 6, 2025 • 11 new comments

• Learn2code | Review of the love2code configs

  #164 commented on Mar 7, 2025 • 0 new comments