iLLM

iLLM is an inspectable inference-runtime learning project. The current runtime is V0.0 Bootstrap Lab: a CPU-only deterministic toy scheduler with typed events, capture, R0 trace playback, and R1 control-state replay. It intentionally contains no tokenizer, tensor, model, GPU, KV cache, or HTTP server yet.

V0.0 quickstart

Use any supported system Python 3.11+ only for the first command. All project commands after bootstrap are resolved from the finalized toolchain manifest and do not depend on shell activation or global PATH entries.

python scripts/bootstrap.py --profile dev --locked --write-manifest out/v0.0/toolchain.json --write-manifest-digest out/v0.0/toolchain.sha256
python scripts/in_env.py --manifest out/v0.0/toolchain.json --expected-manifest-digest-file out/v0.0/toolchain.sha256 -- python scripts/check_quality.py --output out/v0.0/quality.json
python scripts/in_env.py --manifest out/v0.0/toolchain.json --expected-manifest-digest-file out/v0.0/toolchain.sha256 -- python -m pytest tests/cpu --junitxml out/v0.0/junit.xml
python scripts/in_env.py --manifest out/v0.0/toolchain.json --expected-manifest-digest-file out/v0.0/toolchain.sha256 -- python -m illm demo toy-scheduler --seed 7 --steps 20 --capture out/v0.0
python scripts/in_env.py --manifest out/v0.0/toolchain.json --expected-manifest-digest-file out/v0.0/toolchain.sha256 -- python -m illm replay out/v0.0 --mode trace
python scripts/in_env.py --manifest out/v0.0/toolchain.json --expected-manifest-digest-file out/v0.0/toolchain.sha256 -- python -m illm replay out/v0.0 --mode state --assert-state-digest

The normative requirements are in docs/specs/versions/v0.0-bootstrap-lab.md. Later version documents are plans, not implemented capabilities.

Current qualification

V0.0 is an implemented release candidate with local functionally_qualified evidence: CPU, Windows, and Linux profiles pass. The Linux clean-room profile is produced in WSL Ubuntu and merged only after its sealed project-root, bootstrap, lock, environment, and evidence digests match the Windows-built release candidate. GitHub Actions independently repeats the Windows/Linux matrix after each push. The quality gate includes offline mypy --strict, and the runner rechecks the exact locked distribution set before importing iLLM. See docs/implementation/v0.0-status.md and the generated release/v0.0.0/manifest.json.