8 Pull requests merged by 6 people

• Added evaluation config under llama3_1 dir, updated registry

  #2763 merged Jun 2, 2025

• Update valid torch nightly backends

  #2774 merged Jun 1, 2025

• Improve torch.compile log message in compile_model

  #2777 merged Jun 1, 2025

• [compile] Fix graphbreaks in moe split; scale_grad

  #2771 merged May 30, 2025

• Remove code_llama2

  #2770 merged May 30, 2025

• 1M+ context length (context parallel integration)

  #2668 merged May 30, 2025

• [compile] Fix compile producing NaNs

  #2765 merged May 28, 2025

• Update yaml configs. Add validation dataset

  #2608 merged May 28, 2025

4 Pull requests opened by 4 people

• [VERY WIP] DSV3

  #2764 opened May 27, 2025

• [NOT FOR REVIEW] Full knowledge distillation recipe TP + FP8

  #2767 opened May 28, 2025

• Fix typing in _grad_scaler.py for nightly compatibility

  #2778 opened Jun 2, 2025

• Full lora recipy simplification

  #2780 opened Jun 2, 2025

4 Issues closed by 3 people

• Improve `compile_model` logging output

  #2717 closed Jun 1, 2025

• Could you recommend evaluation benchmark for LLM instruction-tuning? (with alpaca or slimorca dataset)

  #2740 closed May 30, 2025

• Deprecate code_llama2

  #2768 closed May 30, 2025

• Generation does not work

  #2769 closed May 29, 2025

4 Issues opened by 4 people

• Proposal: reuse methods in recipes

  #2779 opened Jun 2, 2025

• Loss is `nan` while training LLama4 via LoRA using torchtune

  #2776 opened Jun 1, 2025

• Torchtune dataset cache memory error despite providing dataset file path

  #2775 opened May 30, 2025

• pip install torchtune , version is 0.61 which does not contrain qwen3

  #2772 opened May 30, 2025

16 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.

• [llama4] use grouped_mm in moe for sm90

  #2755 commented on Jun 2, 2025 • 17 new comments

• I'm experimenting unexpected loss values when running old recipes on the same datasets

  #2746 commented on May 27, 2025 • 0 new comments

• Running into a torch.distributed.elastic.rendezvous.api.RendezvousConnectionError: The connection to the C10d store has failed.

  #2676 commented on May 28, 2025 • 0 new comments

• TP + FP8 - NotImplementedError for certain operations

  #2629 commented on May 30, 2025 • 0 new comments

• TP + FP8 + Compile: metadata error

  #2682 commented on May 30, 2025 • 0 new comments

• `torchtune` fails to start when the checkpointer dir does not contain the initial model weights

  #2759 commented on May 30, 2025 • 0 new comments

• Feature request: support different input/output formats in the same recipe

  #2732 commented on May 30, 2025 • 0 new comments

• Perf / accuracy metrics comparison with Nemo for SFT / reasoning distillation scenario

  #2760 commented on May 30, 2025 • 0 new comments

• Implement step based checkpointing

  #2384 commented on Jun 2, 2025 • 0 new comments

• GRPO LoRA Single Device

  #2467 commented on May 31, 2025 • 0 new comments

• [WIP] Gemma3 support.

  #2485 commented on May 28, 2025 • 0 new comments

• Adding EOS Tokens to Qwen Models

  #2512 commented on May 27, 2025 • 0 new comments

• Eval chat template

  #2574 commented on May 28, 2025 • 0 new comments

• [WIP] HuggingFaceModelTokenizer

  #2723 commented on Jun 2, 2025 • 0 new comments

• Add feature ligerceloss

  #2741 commented on May 30, 2025 • 0 new comments

• Add `LRScheduler.state_dict()` to checkpoints

  #2762 commented on Jun 2, 2025 • 0 new comments