Crisol

Zero-bloat heterogeneous compute dispatcher and benchmark runner.

Kubernetes, Slurm, and Ray exist to herd a datacenter. Crisol exists for the lab that actually exists: a laptop, a tower with an RTX, a spare headless box — 2 to 5 machines, not 500. One command, one seed, one honest JSONL ledger per run.

$ crisol status
local-cpu: backend=cpu reachable — native win32; Python 3.12.8

$ crisol jobs
cpu-smoke: backends=cpu default_seed=1729
  Stdlib-only CPU matmul smoke benchmark with honest metrics (no torch required).

$ crisol run cpu-smoke
==================================================
  Job:       cpu-smoke
  Node:      local-cpu (cpu)
  Seed:      1729
  Status:    completed (exit 0)
  Elapsed:   0.8s
  Ledger:    ...\results\2026-08-18.jsonl
==================================================

Why Crisol

• The pain. Setting up Ray/Slurm/K8s for a handful of local boxes is a nightmare of daemons, config, and wasted RAM. Hand-rolled SSH scripts are fragile, produce no traceability, and hide real numbers.
• The reproducibility crisis. People say "my model flies" while comparing different libraries, dimensions, and no warmup. Crisol pins the seed (default 1729), the dimensions, and the warmup so numbers are comparable.
• The audit. Every run appends an immutable JSONL record: run id, job, node, backend, status, exit code, elapsed, seed, and full stdout/stderr.

Names

Install

pip install crisol

Then, from any directory:

# Prefer this if PATH still points at the house lab CLI:
py -m crisol doctor
crisol doctor     # validate python, torch/CUDA, results ledger
crisol status     # probe reachable nodes on this machine
crisol jobs       # list registered jobs
crisol run cpu-smoke    # dispatch a smoke benchmark
crisol run cpu-smoke --seed 42   # override the deterministic seed

cpu-smoke needs no torch — stdlib only. If torch is installed, crisol jobs also lists torch-smoke, which runs a real FP16 matmul on CUDA when available and reports honest TFLOPS.

What you get

• crisol status — probes reachable nodes (local CPU; optional local torch surface with CUDA).
• crisol jobs — lists the registered dispatch contract, including each job's default seed.
• crisol run <job> [--seed N] — dispatches the job as a subprocess and writes a JSONL run record to results/.
• crisol doctor — validates your Python, torch/CUDA availability, and that the results ledger is writable.

The JSONL audit ledger

Each run appends a single line to results/<YYYY-MM-DD>.jsonl:

{"backend": "cpu", "elapsed_s": 0.8, "exit_code": 0, "job": "cpu-smoke",
 "node": "local-cpu", "run_id": "...", "seed": 1729, "status": "completed",
 "stdout": "...", "stderr": "", "timed_out": false, "ts": "..."}

Workloads emit Crisol protocol JSONL on stdout (lifecycle, progress, metrics, result, error) — see docs/PROTOCOL.md.

Roadmap

• More registered hardware benchmarks (FP16 matmul, attention, tok/s).
• SSH transport for the 2-5 box lab — remote dispatch with the same JSONL contract and zero daemons.
• Optional web dashboard for comparing runs.

License

MIT. See LICENSE.

Crisol is the public launch pad of Seedsource Novum. It is deliberately independent of any proprietary physics engine: a clean protocol + metrics + JSONL dispatcher that any lab can adopt.