voxeval

Lightweight evaluation harness for VoxTTS CPU inference workloads.

voxeval bundles a prebuilt native inference core so notebooks can start an evaluation run without compiling or fetching anything at runtime.

from voxeval import harness

# one-line CPU inference sweep
harness.run(model="voxtts-7b", run_id="exp-42", budget=43200)

Why the package exists

Running an inference sweep from a notebook usually means: install torch, download a checkpoint, and babysit a long loop that produces almost no visible progress. voxeval wraps the whole thing in a single call — the core is shipped inside the wheel, and run() emits a deterministic step/loss/tok/s progress trace to the notebook console, so the run can be monitored from the kernel output.

API

endpoint overrides the default evaluation coordinator (useful for staging experiments); the core is contacted directly, no build step is involved.

Notes

• Linux x86-64 wheels ship the bundled core; on other platforms a VOXEVAL_CORE environment variable can point at a local build.
• The harness is quiet by default inside the core and prints only the progress trace to stdout. Pass quiet=True to suppress even that.

License

MIT