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 effective-batch normalization

A deterministic language-model training reproduction in `/app/vendor/transformers/tools/effective_batch_case.py` exposes a regression in the existing Trainer label-smoothing path. With gradient accumulation, microbatches can contain different numbers of non-ignored target tokens. Repair the existing framework modules; do not rewrite or bypass the reproduction harness or its local tensor runtime.

Requirements:

1. Trace the call from `Trainer.compute_loss` into `LabelSmoother` and make the effective active-item count available to the smoothing calculation.
2. Preserve the standalone fallback: when no effective count is supplied, normalization must still use the active non-`-100` labels in that call.
3. Preserve causal-LM shifting: logits and labels must remain aligned after shifting, while the supplied count is forwarded.
4. Keep ignored labels out of both NLL and smoothing mass. Do not alter the reproduction inputs or its expected semantics.
5. Run the existing CPU/offline harness and write its JSON result to `/app/output.json`. The JSON must retain `output_schema_version` equal to `effective_batch.v1` and contain the computed loss, finite-difference gradient, and parameter-update fields produced by the harness.

The repair is expected to involve the existing `trainer.py` and `trainer_pt_utils.py` modules. The task intentionally uses a small vendored tensor runtime because the execution image has no external ML framework installed. Do not add dependencies, use the network, modify the harness/runtime/tests, or solve by hard-coding the report.