Repository navigation
Studio: duplicate a past training run into a new Configure draft - #12979
Conversation
|
Duplicate is available from the Recents run menu and from the Training config popover of a past run. It loads that run's saved settings into a fresh Configure draft and switches to Configure without starting training. Checked against a seeded finished run on the merge base Optional, not blocking: Limits: Studio ran CPU-only with the Train view enabled through a stubbed capability flag. Windows and macOS were not run natively. Fork CI is still running: Etherll#186
|
|
Yay |
Adds a Duplicate action to the sidebar run menu and the training config popover. It loads the run's saved config into a fresh Configure draft and switches to the Configure tab without starting training. Credentials, remote-code approval, output paths and resume state are not copied. Closes unslothai#12977
The sidebar and progress section imported the hook by its internal path, which the no-restricted-imports rule rejects. Export it from @/features/training like the other training hooks.
bcc14e5 to
75eb68e
Compare
|
Sorry I was removing the Co authored by claude 😭 |
|
LGTM! Thanks for the PR ! |
…'s token after #13005 (#13022) * Hub token identity tests: let the llama-backend stub take the token #12979 now passes #12979 (184e1fc) made `_llama_backend_active(model, token)` so the embedding resolver can gate the llama-server path on the caller's Hub token, and `_resolve_embedding_model_plan` now calls it with both. The two stub tables in `test_hub_token_caller_identity.py` still hand it a one-argument lambda, so every test built on `_st_resolver` / `_gguf_resolver` fails with `TypeError: <lambda>() takes 1 positional argument but 2 were given`: seven cases, red on Backend CI `(Python 3.13, a-k)` for 184e1fc and 04013e9. The stubs now accept the token and ignore it, which is what they did with the model. `tests/test_hub_token_caller_identity.py`: 197 passed. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_012DRiTpjxAymRtv9p8Xr2kd * Pin that the embedding resolver asks the backend probe with the caller's token #13005 made _resolve_embedding_model_plan pass the caller's Hub token to _llama_backend_active, because whether llama-server or sentence-transformers serves a private repo depends on files only that token can read. The stubs fixed above made the suite pass again but nothing checked the token still arrives; this test fails if the resolver drops it. * Shorten the new test's docstring --------- Co-authored-by: Claude Fable 5.1 <[email protected]> Co-authored-by: Daniel Han <[email protected]>
Summary
Closes #12977.
Adds a Duplicate action for past training runs in Unsloth Studio. It loads the run's saved config into a fresh Configure draft and switches to the Configure tab. It never starts training, so the user can review and edit before pressing Start Training.
Entry points:
⋮menu in the sidebar Recents, next to Rename and DeleteScreenshots
What is copied, and what is not
_sanitize_db_configalready strips HF/W&B tokens and S3 keys, and the draft clears the W&B token.trust_remote_codeand its fingerprint).Duplicating replaces the current Configure draft.
🤖 Notice: Made with Claude and Astra but I kept an eye on them