Summary

• Keep the public path on the existing evaluate_and_print_results() helper: it still performs the same logging/printing side effects and now returns the averaged loss dictionary when evaluation completes.
• Remove the separate evaluate_model() wrapper from this branch so there is no second eval entry point.
• Add examples/evaluation/pure_eval.py, a runnable minimal Bridge example that builds a small GPT model, mock data iterator, and calls evaluate_and_print_results() directly.
• Keep focused mock unit coverage for the existing helper's return-value behavior.

Usage

uv run python -m torch.distributed.run --nproc_per_node=2 examples/evaluation/pure_eval.py --tp-size 2

Programmatic callers can use the existing helper directly:

losses = evaluate_and_print_results(...)

The example uses the standard Bridge setup path to create the model, mock dataset iterator, and process groups, then calls evaluate_and_print_results(). Workflows that already have a Megatron model, ConfigContainer, and iterator can use the same existing helper call.

Tests / Validation

• ruff check src/megatron/bridge/training/eval.py tests/unit_tests/training/test_eval.py examples/evaluation/pure_eval.py
• ruff format --check src/megatron/bridge/training/eval.py tests/unit_tests/training/test_eval.py examples/evaluation/pure_eval.py
• python3 -m py_compile src/megatron/bridge/training/eval.py tests/unit_tests/training/test_eval.py examples/evaluation/pure_eval.py
• git diff --check origin/main...HEAD
• pre-commit run --all-files
• Internal 2-GPU validation on pushed head e1528cc:

uv run --no-sync python -m torch.distributed.run --nproc_per_node=2 \
  examples/evaluation/pure_eval.py --tp-size 2 --eval-iters 1 \
  --seq-length 64 --hidden-size 64 --num-attention-heads 4 \
  --global-batch-size 2 --micro-batch-size 1

Result: passed. The run initialized tensor model parallel size 2, evaluated one iteration, and returned {'lm loss': 10.390985488891602}.

Not run as passing local evidence:

• uv run pre-commit run --all-files is blocked on this workstation before hooks run because the pinned nvidia-resiliency-ext==0.6.0 wheel is not available for the local Linux platform baseline.
• Runtime pytest was not used as authoritative local evidence; runtime validation was done with the 2-GPU example smoke above.

Notes

• The example intentionally keeps using the existing setup path, which still constructs optimizer/scheduler objects because current setup/config validation expects them. This PR does not refactor those broader requirements.

Fixes #2033