Prepared artifact for nathanrs/detllm: an inference pipeline where the entire forward pass is exact integer arithmetic, producing bit-identical logits across runs, batch compositions, prefill/decode splits, and hardware backends (NVIDIA tensor cores vs. pure-CPU reference).

Derived from Qwen/Qwen3-0.6B (Apache-2.0) by one-time offline preparation (per-channel int8 weight quantization with folded norms and analytic smoothing, per-row-quantized embedding, int16 RoPE tables, dyadic scale constants). WikiText2 perplexity 20.72 vs the fp16 baseline's 20.95.

Integrity: runs on different machines are only comparable when sha256(model.safetensors) matches — the repo's demo prints it on startup.

Usage: