test-zentorch-unit

name	test-zentorch-unit
description	Run zentorch tests (unit, op, model, miscellaneous, export, vLLM, LLM, or pre-trained model tests) with the required ZenDNN cache env vars. Use when the user asks to run or discover zentorch tests.

Run zentorch tests

When the user asks to run tests, follow this skill.

Agent action: Run the mapped command (adjust scope as needed):

.claude/skills/test-zentorch-unit/scripts/test.sh                     # unit tests (default)
.claude/skills/test-zentorch-unit/scripts/test.sh all
.claude/skills/test-zentorch-unit/scripts/test.sh op_tests
.claude/skills/test-zentorch-unit/scripts/test.sh model_tests
.claude/skills/test-zentorch-unit/scripts/test.sh miscellaneous_tests
.claude/skills/test-zentorch-unit/scripts/test.sh export_tests
.claude/skills/test-zentorch-unit/scripts/test.sh vllm_tests
.claude/skills/test-zentorch-unit/scripts/test.sh llm
.claude/skills/test-zentorch-unit/scripts/test.sh pre_trained
.claude/skills/test-zentorch-unit/scripts/test.sh test/unittests/op_tests/test_bmm.py

The script validates the requested scope before any package checks or installation, verifies that zentorch is installed, sets the required env vars, installs test deps, and runs the scope. Use the manual commands below only if the script fails or a -k / -p filter is needed that the script does not support.

See zentorch-test-flow.md for the test workflow.

---

Environment

Requires an activated, non-base Python environment with zentorch installed — the same environment used across the other zentorch skills. Confirm what is active:

echo "${VIRTUAL_ENV:-${CONDA_DEFAULT_ENV:-none}}"

If zentorch is not built/installed yet, follow the build-zentorch-setup-env or build-zentorch-from-source skill first.

---

1. Validate what to run

Match the user's request to the right command. If the user just says "run tests" with no specifics, run all unit tests:

python -m unittest discover -s ./test/unittests

The script rejects an unknown scope or nonexistent file immediately, before checking zentorch or running the dependency installer. A direct test file is run with python -m unittest <file>; it does not use discovery.

2. Verify zentorch and set required environment variables

The script first verifies import zentorch. Then disable ZenDNN caching before running any tests:

python -c "import zentorch"
export ZENDNNL_MATMUL_WEIGHT_CACHE=0
export ZENDNNL_ZP_COMP_CACHE=0
export ZENDNNL_ENABLE_POSTOP_CACHE=0

3. Install test dependencies

python test/install_requirements.py

Installs transformers, expecttest, parameterized, hypothesis, and deprecated, plus the torchvision and torchao versions matching the installed PyTorch.

Test categories

Every category is discovered the same way (python -m unittest discover -s <path>):

User says something like	Scope	Path
"run all tests"	all	./test
"run unit tests" / "unittests"	unittests	./test/unittests
"run op tests"	op_tests	./test/unittests/op_tests
"run model tests"	model_tests	./test/unittests/model_tests
"run miscellaneous tests"	miscellaneous_tests	./test/unittests/miscellaneous_tests
"run export tests"	export_tests	./test/unittests/export_tests
"run vllm tests"	vllm_tests	./test/unittests/vllm_tests
"run llm tests"	llm	./test/llm_tests
"run pre-trained / pretrained"	pre_trained	./test/pre_trained_model_tests

If the user names a specific op or feature

Find the matching test file first, then run it:

find test -name 'test_*.py' | grep -i "<keyword>"
python -m unittest <path>

Examples (verified paths):

User says	Run
"run bmm test"	python -m unittest test/unittests/op_tests/test_bmm.py
"run mm test"	python -m unittest test/unittests/op_tests/test_mm.py
"run linear tests"	python -m unittest discover -s ./test/unittests -p "test_linear*"
"run woq tests"	python -m unittest discover -s ./test/unittests -k "woq"
"run sdpa test"	python -m unittest test/unittests/model_tests/test_sdpa.py
"run fused moe test"	python -m unittest test/unittests/model_tests/test_fused_moe.py
"run version test"	python -m unittest test/unittests/miscellaneous_tests/test_zentorch_version.py
"run bert test"	python -m unittest test/pre_trained_model_tests/test_bert.py

If multiple files match, show the matches and ask which to run, or use -k / -p filters to run all of them.

Generated export-test packages

The underlying export tests may write model.pt2 and model_z.pt2 in the repository root. The test skill intentionally does not delete or restore files by name because it cannot determine whether they predated the run. Check those paths before running export tests, and clean up only artifacts known to belong to that run.

---

Test directory structure

test/
  unittests/
    op_tests/              # Individual op tests (bmm, mm, embedding, qlinear, rms_norm, ...)
    model_tests/           # Custom model fusion/graph tests (sdpa, fused_moe, pattern_matcher, ...)
    miscellaneous_tests/   # Non-op tests (version, device/isa checks)
    export_tests/          # torch.export tests
    vllm_tests/            # vLLM integration tests
  llm_tests/               # Shared LLM test utilities (llm_utils.py)
  pre_trained_model_tests/ # Real model tests (bert, cnn)
  zentorch_test_utils.py   # Shared test utilities
  install_requirements.py  # Test dependency installer

Test file quick reference

op_tests/

test_addmm, test_baddbmm, test_bmm, test_dynamic_qlinear, test_embedding, test_embedding_bag, test_embeg_pack_weight, test_group_matmul, test_horizontal_embedding_bag_group, test_horizontal_embedding_group, test_linear, test_matmul_direct, test_matmul_tanh, test_mm, test_mm_silu, test_mm_silu_mul, test_prepare_4d_causal_attention_mask, test_qlinear, test_qlinear_eltwise, test_qlinear_mul_add, test_quant_embedding_bag, test_rms_norm, test_woq_linear, test_woq_linear_asymmetric

model_tests/

Custom models exercising fusion/graph patterns: addmm/mm unary + binary fusions, linear unary/binary, qkv fusion, sdpa, fused MoE, horizontal embedding groups, qlinear reorder/eltwise, woq fusions, and the pattern matcher.

miscellaneous_tests/

File	What it tests
test_zentorch_version.py	zentorch version string
test_avx512_device.py	AVX-512 device/ISA detection
test_bf16_device.py	bf16 device support
test_fp16_device.py	fp16 device support

export_tests/

File	What it tests
test_exp.py	torch.export support

vllm_tests/

File	What it tests
test_vllm_plugin.py	vLLM plugin integration

pre_trained_model_tests/

File	What it tests
test_bert.py	BERT model inference
test_cnn.py	CNN model inference