You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs(examples): aLora/102_example.py fails with AloraActivationError (stembolts io.yaml has no instruction) #1790
Following the aLoRA README, docs/examples/aLora/102_example.py fails on main with:
AloraActivationError: aLoRA 'stembolts' (base model 'granite-3.3-2b-instruct', qualified
name 'stembolts_alora') never activates for this call: its declared invocation sequence
(49152, 1553, 81, 34775, 49153) is absent from the assembled prompt
The example is skip-marked, so CI never runs it. It is the only example of a custom, non-catalog aLoRA adapter, and docs/examples/aLora/README.md currently describes stembolts_intrinsic.py (the helper it drives) as a runnable version.
Cause
The stemboltsio.yaml has no instruction: field. The aLoRA's alora_invocation_tokens decode to <|start_of_role|>check_requirement<|end_of_role|>, which therefore never reaches the prompt, so the activation guard added in #1684 correctly refuses to run it. Catalog adapters get the sequence from an instruction: field in their io.yaml (for example requirement-check).
docs/examples/aLora/io.yaml is byte-identical to the one on nfulton/stembolts, and m alora train defaults its invocation prompt to the same <|start_of_role|>check_requirement<|end_of_role|>. So the README's own train and upload flow produces an adapter that cannot activate.
the example runs and returns {"defective_part": "Piston Rings", "diag_likelihood": 0.9, ...}. Without it, the error above, identically at the PR #1765 head and on current main (972fa64).
Reproduce (needs uv sync --extra hf):
printf'The engine sputters at high RPM after sitting overnight.\nOil seepage around the piston rings.\n' \
| uv run python docs/examples/aLora/102_example.py
To decide
Add the instruction: field to docs/examples/aLora/io.yaml and the README flow.
Update nfulton/stembolts on Hugging Face (not something a PR here can change).
Whether m alora train or m alora upload --io-yaml should add the field automatically, so users training their own adapter don't hit this.
Whether the example should be un-skipped or covered by a test once it works.
Activity