Separate compilation, execution and tracing evidence

First-call latency combines tracing, compilation and execution. Repeated calls can have different costs. This small CPU workload checks the numerical output, synchronizes each measurement, then calls the same jitted function at two matrix shapes. It retains a JAXPR and a JSON report. It does not measure device memory or provide a profiler trace.

Before you start

Complete performance-01 through performance-05. Know why asynchronous JAX work must be synchronized before timing. Use the CPU environment from requirements-cpu.txt, and edit projects/performance-audit/starter/model.py.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-cpu.txt
python3 projects/performance-audit/tests/check.py --implementation starter --stage 1
python3 projects/performance-audit/tests/check.py --implementation starter --stage 2
python3 projects/performance-audit/tests/check.py --implementation starter --stage 3 --evidence-dir runs/performance-audit
python3 projects/performance-audit/tests/check.py --implementation starter --stage 4

On Windows PowerShell, activate .venv\Scripts\Activate.ps1 and use python instead of python3.

Stage 1 compares the result for the identity matrix against NumPy. Since $I I=I$, the diagonal entries of $\tanh(I)$ are about 0.761594 and the off-diagonal entries are zero. The first timing includes the first call; later timings are synchronized steady calls. Do not interpret one machine's microseconds as a general performance result.

Stage 2 invokes the same compiled function at shapes 16 by 16 and 32 by 32. The Python trace counter is read after both calls and before JAXPR capture, so it counts shape-driven function traces only. Expect two. The separate make_jaxpr artifact-generation call is excluded from this count.

Stage 3 writes report.json and workload.jaxpr.txt under runs/performance-audit/. The report explicitly says memory_measured: false. Stage 4 integrates the checks. Passing output begins PASS performance-audit stage N.

This records CPU first-call and repeated-call observations plus a symbolic program. It is not a hardware profile, a memory measurement, or TPU qualification. Reference command: python3 projects/performance-audit/tests/check.py --implementation solution --stage 4.