github.com/jakubprejzner
Code that looks the same when nobody is watching.
Stack: Python, PyTorch, quantization (QuIP#, AQLM, GPTQ), FastAPI, Docker, Prometheus/Grafana. Public repositories, MIT licensed.
Serving models in production is not model.predict() behind Flask — it is a model registry with lazy initialization, A/B tests and full observability.source: ml-serve (MIT)
The high end of the workshop. Six techniques, six proofs.
everything below is measured and signed with a commitWeights encoded as vectors on the E8 lattice rather than scalar-wise — Bielik’s 11 billion parameters packed into ~2.4 bits per weight without losing the Polish.
22 GB → 3.26 GB93.2% FP16 quality retentionarXiv · Table 4Second-order information computed once on a Polish corpus (CulturaX-PL) and shared across six quantization methods — 200 files published for the community to reuse.
25.6 GB / 40 minH200 · 18× speedup · $4HF dataset · Sec. 3.267 tasks, zero randomness: greedy decoding verified bit for bit across model load cycles, an oracle replays every result from stored trajectories, and ordering traps separate reading the results from guessing them.
20/20 runsreproduced exactlyPolAgentBench · draftTasks constructed so that forecast shifts cancel exactly — the result is invariant to intermediate rounding. The artifact we caught moved the apparent threshold by a full bit; after correction: zero contaminated slots.
14/96 → 0contaminated arithmetic slotsPolAgentBench · Sec. 3.5QTIP fine-tuning after quantization, with a full-precision teacher — peak usage 112 GB of VRAM on an H200; it recovers the generation quality that quantization alone takes away.
+2.02 pp GEN112 GB VRAM · 2 h · $8arXiv · Sec. 3.6 / Table 10A model registry with lazy initialization, A/B across variants, rate limits, Prometheus metrics and a ready Grafana dashboard — onedocker compose up and it is all running.