tasks/candidate-1866-ml-inference/instruction.md

You are an agent, your current working directory is /app.

You can use the tools available to you to interact with the computer to assist the user in completing tasks.

Repair the frozen Hugging Face Transformers source tree under /app/vendor/transformers. A pretrained inference path currently loses the adapter keyword-argument state at the boundary between adapter discovery and from_pretrained.

Complete the following workflow:

1. Inspect the adapter integration helper and its caller in the pretrained model loader. Repair both existing production modules so their return/unpack contract remains synchronized.

2. Preserve adapter keyword arguments through local adapter discovery. A private _adapter_model_path hint must be consumed by discovery and must not be forwarded later, while ordinary adapter options remain unchanged.

3. Preserve offline cache behavior: local_files_only, revision, subfolder, and resolved commit metadata belong to discovery and must not be mixed into the adapter kwargs passed to the later adapter load.

4. Keep early-return behavior valid when PEFT is unavailable or the model path is absent, including the caller-visible adapter kwargs value.

5. Run the included integration workflow and write /app/output.json:

   python3 /app/tools/run_adapter_workflow.py \
     --cases /app/fixtures/adapter_cases.json \
     --output /app/output.json
   

The output must use schema version transformers.adapter_contract.v1, include one result for every fixture case, describe the repaired caller contract, and set every reported check to true. Compile the two modified modules before finishing. Do not modify the fixture, runner, unrelated vendored files, or add a replacement evaluator. Runtime network access and package installation are not available.