**inferlast** is a CPU-first inference optimizer: prove a real optimization win on a
plain laptop CPU before you ever spend on a GPU.
What's in this release
- **`trustcheck` — the false-win catcher.** Audits whether a benchmark "win" can be
believed. It caught *this repo's own* recorded evidence: the same INT8-vs-fp32
config reported `3.0x faster` in one session and `0.65x slower` in another.
`trustcheck` says **FALSE — that's noise** (CI [-1.78x, 5.45x] straddles 1.0x).
- **`gpucheck` — do you even need a GPU?** A CPU-only decision rule that returns
`GPU-warranted / CPU-suffices / insufficient-data`, and refuses to guess when the
data isn't there.
- **Profiler** for the CPU bottleneck (overhead-bound, not weight-bound, on tiny
models).
- **Auto-quantization** with a robust logit-geometry quality metric — it said
*KEEP fp32* on our machine because INT8 shifted the answers. That refusal is the
feature.
- **Auto-batching** latency/throughput sweep.
Honesty, published
Every number here is measured on a 2019 Intel MacBook Pro (CPU-only, 16 GB), not
claimed — including the failures. 55 tests, no model downloads.
The one-line thesis
Fine-tuning-free, GPU-free: inferlast is the decision + trust layer for CPU-first
inference. The falsifiable claim it's built on lives in [docs/RESEARCH-SPEC.md](docs/RESEARCH-SPEC.md).
Try it
bash
pip install -r requirements.txt
python scripts/run_all.py --model Qwen/Qwen2.5-0.5B-Instruct