Update torchtune component paths from utils to training in torchtune notebook

[johndmulhausen]

What

Running the config generated by colabs/torchtune/torchtune_and_wandb.ipynb fails with:

AttributeError: module 'torchtune.utils' has no attribute 'metric_logging'

torchtune relocated these APIs out of torchtune.utils (and one out of torchtune.modules) into torchtune.training. This PR updates the stale _component_ paths in the notebook's markdown YAML snippet and the %%writefile mistral_wandb_lora.yaml cell:

Old path	New path
torchtune.utils.metric_logging.WandBLogger	torchtune.training.metric_logging.WandBLogger
torchtune.utils.FullModelHFCheckpointer	torchtune.training.FullModelHFCheckpointer
torchtune.utils.profiler	torchtune.training.setup_torch_profiler
torchtune.modules.get_cosine_schedule_with_warmup	torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup

The lr_scheduler and profiler paths match the ones used in torchtune's own bundled recipes/configs/mistral/7B_lora_single_device.yaml. The profiler path is also the only value the lora_finetune_single_device recipe accepts (it asserts _component_ == torchtune.training.setup_torch_profiler). Inline # <--- comments and all other YAML content are unchanged; the diff is 5 lines.

Verification (against torchtune 0.6.1, torch 2.8.0)

• from torchtune.training.metric_logging import WandBLogger imports cleanly.
• Extracted the full YAML from the edited %%writefile cell, walked every _component_: value, and resolved each with torchtune.config._utils._get_component_from_path (the exact code path that crashed in the issue) — all 9 components resolve:

OK   tokenizer:     torchtune.models.mistral.mistral_tokenizer
OK   dataset:       torchtune.datasets.alpaca_dataset
OK   model:         torchtune.models.mistral.lora_mistral_7b
OK   checkpointer:  torchtune.training.FullModelHFCheckpointer
OK   optimizer:     torch.optim.AdamW
OK   lr_scheduler:  torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
OK   loss:          torch.nn.CrossEntropyLoss
OK   metric_logger: torchtune.training.metric_logging.WandBLogger
OK   profiler:      torchtune.training.setup_torch_profiler

• tune validate mistral_wandb_lora.yaml on the extracted config reports Config is well-formed!
• The lora_finetune_single_device recipe used by the notebook still exists in tune ls.
• nbformat.validate passes; no other torchtune.utils references remain in the notebook.

Notes

• The notebook installs torchtune from an unpinned git clone --depth 1 of pytorch/torchtune main, so the hardcoded component paths track a moving upstream — the same drift that caused this breakage. Left as-is to keep the diff minimal, but pinning a release could prevent a repeat.
• A full training run was not executed (needs the Mistral-7B checkpoint download and a GPU); verification covers config parsing and component resolution, which is where the reported crash occurred.

Fixes #579

🤖 Generated with Claude Code

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[Copilot]

Pull request overview

Updates the torchtune notebook’s YAML _component_ references to match torchtune’s newer module layout (torchtune.training), fixing runtime config instantiation failures caused by stale torchtune.utils / torchtune.modules paths.

Changes:

• Update W&B metric logger component path to torchtune.training.metric_logging.WandBLogger (in both the markdown snippet and the written YAML cell).
• Update checkpointer component path to torchtune.training.FullModelHFCheckpointer.
• Update LR scheduler and profiler component paths to torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup and torchtune.training.setup_torch_profiler.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.