Search results

How to separately load a local llama8B and a self-trained LoRA adapter during distillation?

Status: Open.

#2520 In pytorch/torchtune;

· whale2133 opened on Mar 21, 2025

Use foreach compilable scale/grad / clip_grad

Status: Open.

#2517 In pytorch/torchtune;

· IvanKobzarev opened on Mar 19, 2025

Add torch.compile to optimizer.step()

Status: Open.

#2516 In pytorch/torchtune;

· IvanKobzarev opened on Mar 19, 2025

Unnecessarily scaling gradients when gradient_accumulation_steps is 1

Status: Open.

#2515 In pytorch/torchtune;

· shunting314 opened on Mar 19, 2025

[Model] Support for Mistral Small 3.1

Status: Open.

#2508 In pytorch/torchtune;

· maximegmd opened on Mar 17, 2025

Can we decouple the data preprocssing/tokenization step from the fine-tuning phase?

Status: Open.

#2497 In pytorch/torchtune;

· Electronic-Waste opened on Mar 13, 2025

cross-attention in llama 3.2 vision has is_causal=False

Status: Open.

#2493 In pytorch/torchtune;

· mmehdig opened on Mar 13, 2025

Consolidate tok_encode logic between _LLMEvalWrapper and _VLMEvalWrapper

Status: Open.

#2488 In pytorch/torchtune;

· pbontrager opened on Mar 12, 2025

Gemma3

Status: Open.

#2484 In pytorch/torchtune;

· krammnic opened on Mar 12, 2025

recursive_reshard

Status: Open.

#2483 In pytorch/torchtune;

· caiqi opened on Mar 12, 2025

Add add_end_token to the Qwen Models

bugSomething isn't workingSomething isn't working

community help wantedWe would love the community's help completing this issueWe would love the community's help completing this issue

good first issueGood for newcomersGood for newcomers

Status: Open.

#2481 In pytorch/torchtune;

· pbontrager opened on Mar 11, 2025

Add add_end_token to Phi tokenizers

bugSomething isn't workingSomething isn't working

community help wantedWe would love the community's help completing this issueWe would love the community's help completing this issue

good first issueGood for newcomersGood for newcomers

Status: Open.

#2480 In pytorch/torchtune;

· pbontrager opened on Mar 11, 2025