Testing¶

RL-Kernel uses focused tests for dispatch behavior and operator accuracy.

gtest (operator candidate vs gold)¶

Primary entry for single-operator forward/backward checks against a PyTorch gold path:

python scripts/check_operator.py --op logp --candidate pytorch --device cpu --dtype fp32 \
  --batch 1 --seq 2 --vocab 17

Full usage (register OP_SPECS, build inputs, CLI flags, and the WS1 four-judgment tolerance contract after #267):

gtest usage guide (operator CLI + OP_SPECS + contract; English)

Dispatch Tests¶

python -m pytest rl_engine/tests/test_dispatch.py -v

Operator Accuracy¶

python tests/test_op_accuracy.py

Contract schema / resolver and WS1 C1–C8 CPU gates:

python -m pytest tests/test_tolerance_contract.py tests/test_op_checks.py \
  tests/test_ws1_workload.py tests/test_forward_invariance.py \
  tests/test_gradient_invariance.py tests/test_elementwise_inventory.py \
  tests/test_four_judgment_matrix.py tests/test_operator_inputs.py -q

max_abs_dlogp, approx_kl0, and clipfrac0 are the sole chain-level logprob aggregates; gradient pass/fail uses independent gradient_* verdicts.

CUDA BF16 and Triton-on-CUDA BF16 gtest + C8 --execute run in .github/workflows/ws1-gtest-gpu.yml (RunPod). Local equivalent:

bash ci/run_ws1_gtest.sh

The C8 JSON is written outside the repo (${TMPDIR:-/tmp}/ws1-c8-ci.json unless WS1_C8_JSON is set) so the recorded git provenance is not dirtied by the artifact itself. The GitHub workflow uploads that file as a CI artifact.

Documentation Build¶

pip install -r requirements-docs.txt
mkdocs build --strict -f mkdocs.yaml

Run the documentation build whenever adding a new operator page or changing navigation.