Skip to content

testing.md

Latest commit

 

History

History
149 lines (102 loc) · 10.4 KB

testing.md

File metadata and controls

149 lines (102 loc) · 10.4 KB
Raw

Testing and Benchmarking

Tests and benchmarks are separated by concern: pytest tests/ validates correctness only; pytest benchmarks/ runs profiling only and auto-generates profile_run.log.

Core Abstractions

Class Location Role
WorkloadBase workloads/workload_base.py ABC defining gen_inputs(). Shared base used by both tests and benchmarks; a subclass named for one op also defines that op's ref_program().
FixtureBase workloads/workload_base.py Metaclass-based decorator that applies pytest.mark.parametrize from a PARAMS class attribute or get_params() classmethod.
TestBase tests/test_base.py Inherits WorkloadBase. Declares ref_program() abstract and adds check(). Each op subclasses this for correctness testing.
BenchmarkBase[W] benchmarks/benchmark_base.py Generic ABC parameterized by workload type W (a capability protocol, not WorkloadBase). Subclass implements calculate_flops() and calculate_memory(). Provides profile().
BenchmarkReport benchmarks/benchmark_base.py Static collector -- record() stores results, dump() writes markdown, clear() resets.

Wiring

Workload is defined once; test and benchmark each reference it but do not depend on each other:

  • Workload (workloads/) — WorkloadBase subclass, defines gen_inputs() and, when named for one op, ref_program()
  • Test (tests/ops/) — inherits (Workload, TestBase), adds tolerances; defines ref_program() only on a shape-only workload
  • Benchmark (benchmarks/ops/) — composes workload via BenchmarkBase(workload)

Rules:

  • Fixture usage: both tests and benchmarks can use FixtureBase, but params are usually defined per layer unless intentionally factored into a shared module
  • Dependency direction: benchmark imports workload, never test
  • ref_program locality: the reference lives on the narrowest shared class that names one operator — the workload, unless the workload describes only an input shape

Tests

→ Trust boundary: trust-model.md §Test | Rules: testing-budget.md

Framework: pytest. Location: tests/ops/.

File checklist

  1. Workload class in workloads/ — subclass WorkloadBase, implement gen_inputs() and, when the class is named for one op, ref_program().
  2. Fixture class — subclass FixtureBase, define PARAMS with smoke/full marks.
  3. Test class in tests/ops/test_<op>.py — inherit (MyWorkload, TestBase). Implement ref_program() here only when the workload describes an input shape rather than an op.
  4. Test function — @YourFixture decorated, call test.check(op, *test.gen_inputs()).

Tolerance

  • Use torch.testing.assert_close for floating-point verification:
    • FP16: rtol=1e-3, atol=1e-3
    • BF16: rtol=1.6e-2, atol=1.6e-2
  • Use exact comparison (torch.equal) for non-floating outputs (bool, masks, index tensors).

Coverage rules

  • Tests must cover FP16 and BF16 data types.
  • Tests must parameterize over common shapes (batch size, heads, sequence length).
  • Tests must encode the dtype contract: supported dtypes are covered, unsupported dtypes are rejected, output dtypes are asserted when they differ from input.
  • GPU-dependent tests must run on a real machine with host-visible CUDA devices. Sandbox-only results are not final correctness evidence.

Test case policy

Each parameterized case must serve one of:

  1. Dtype correctness — verify a supported dtype.
  2. Shape coverage — verify a distinct code path (boundary, tile edge, alignment).
  3. Feature coverage — verify a feature flag or mode (causal=True, tune=True).
  4. Regression — reproduce a fixed bug (reference issue/PR in comment).

No performance exploration, autotune sweeps, or duplicate code-path coverage.

Dtype coverage: All supported dtypes must be tested. Smoke: cover each dtype with one typical shape. Full: cross-combinations only when the implementer can name the code path each guards.

Shape coverage: UT shapes target kernel implementation branches, not workload representativeness. Common kernel branch conditions:

  • Tile boundary — shape not divisible by tile size (tail handling)
  • Vectorization alignment — shape not aligned to vector width (scalar fallback)
  • Degenerate dimension — size=1 (broadcast, squeeze paths)
  • Dispatch branch — different shape ranges triggering different kernel variants

The implementer selects the smallest shape that triggers each branch. Do not generate test fixtures from src/tileops/manifest/ workloads.

Growth rules:

  • Each new case must state its purpose (dtype / shape / feature / regression) in a comment or PR description.
  • Over 20 cases per test function: justify which code paths require the count.
  • Prefer a new test function over inflating an existing one when testing genuinely different behavior.

Test node growth detection

scripts/test_node_delta.py compares pytest collected node count (test cases after parametrize expansion) between current branch and main. Always exits 0 (non-blocking).

python scripts/test_node_delta.py                    # auto-detect changed test files
python scripts/test_node_delta.py tests/ops/test_foo.py  # specific files
python scripts/test_node_delta.py --base origin/release   # different base branch
  • No growth on existing files: nothing to report.
  • Growth on existing files: include script output and a one-line justification in PR description.
  • New test files only: no delta to report — follow the policy above.

Testing layers

Layer Responsibility Shape source
UT smoke/full Guard PR correctness Implementer selects based on kernel code paths
Nightly benchmark Performance regression + typical/stress correctness src/tileops/manifest/ workloads
Local dev Performance tuning verification Developer decides ad-hoc

Infrastructure rules

  • Changes to shared test infrastructure (tests/test_base.py, common fixtures, shared comparators) must preserve existing default semantics unless all affected tests are migrated in the same PR.
  • If a PR touches shared test infrastructure, run a broader pytest -m smoke pass before merge.
  • Run full targeted test files for the affected op family on a real GPU before claiming readiness.

Benchmarks

→ Trust boundary: trust-model.md §Benchmark | Rules: benchmark.md

Framework: benchmarks.benchmark_base.BenchmarkBase. Location: benchmarks/ops/.

Execution: pytest benchmarks/ auto-generates profile_run.log (markdown format).

Workloads

Import the op's workload from workloads/. BenchmarkBase[W] is generic over workload type and reads no attribute off it — ManifestBenchmark takes its roofline from op.eval_roofline() — so a workload needs nothing beyond the fields its own benchmark reads.

File checklist

  1. Workload — import the op's class from workloads/. If the op has none, add it there first: a benchmark must not author gen_inputs.
  2. Fixture class — use FixtureBase with benchmark-specific PARAMS, or pytest.mark.parametrize directly.
  3. Benchmark class in benchmarks/ops/bench_<op>.py — subclass BenchmarkBase, implement calculate_flops() and calculate_memory() (return None if not applicable).
  4. Benchmark function — @YourFixture decorated, construct workload + benchmark, call inputs = workload.gen_inputs(), then bm.profile(op, *inputs) and BenchmarkReport.record(op, locals(), result, tag="tileops").
  5. Independent baseline — record at least one non-"tileops" baseline (e.g., "torch", "fa3"). Profile the workload's ref_program for the torch baseline; a baseline that is deliberately not the reference (different layout, faster idiom, external library) overrides ref_program in the benchmark and says why. Never import a baseline from tests/.

Metrics

  • Latency (ms)
  • TFLOPS (Tera Floating-point Operations Per Second)
  • DRAM Bandwidth (GB/s)

Reporting rules

  • Numbers must come from a real GPU machine, not a sandbox.
  • Include small, medium, and large representative shapes.
  • Do not cherry-pick favorable shapes; report regressions as-is.
  • Run the targeted correctness suite on the same GPU before reporting benchmark numbers.
  • BenchmarkReport.record() first argument may be the Op instance or a string name; stay consistent within a given benchmark file.
  • calculate_flops() and calculate_memory() should return numeric values when the metric is available; return None only if the metric is not applicable, in which case it will be omitted from the report.
  • Every benchmark must record at least one non-"tileops" baseline. Use existing tags ("baseline", "torch", "fa3", "fla", "triton") and avoid introducing ad-hoc tags without updating downstream consumers.