Matricxon testing

Unit and API tests

.venv/bin/pytest            # unit and API tests with tiny synthetic models, no downloads
.venv/bin/ruff check

Checks against real models

Forward passes and tokenizers are compared with Hugging Face on real models:

.venv/bin/python -m tools.oracle.validate_qwen3          # also validate_qwen2, validate_bert, ...
.venv/bin/python -m tools.oracle.validate_tokenizer

tools/README_oracle.md in the repository explains how to run these on a computer with limited RAM.

tools/manual_chat_check.py, manual_embed_check.py and manual_pull_check.py exercise the real HTTP API end to end.

Long benchmarks

Long benchmarks can overheat a laptop. Use tools/thermal_guard.py to stop them at a safe temperature.