CLI reference

turbopress --version
turbopress {compress,validate} ...

turbopress compress

Quantize every decoder linear of a Hugging Face causal LM and write a certified, self-contained artifact.

turbopress compress <model> [options]

flag	default	description
model	—	Hugging Face model id or local path
--bits	3	bits/weight, 2–6
--n-states	64	trellis states: 4, 8, 16, 64 (16 = faster, 64 = best)
--out	turbopress_out	output directory
--seqlen	512	tokens per eval/calibration sequence
--eval-seqs	8	eval set = eval-seqs × seqlen tokens
--calib-seqs	8	disjoint calibration slice
--batch	2	sequences per forward
--seed	0	RNG seed (rotation signs are stored, so runs are reproducible)
--device	auto	cuda or cpu
--no-self-test	off	skip reloading the artifact to verify it round-trips

turbopress validate

Measure KL(reference ‖ candidate), top-1 next-token agreement, and perplexity of both models on held-out text. The two models must share a tokenizer/vocab.

turbopress validate <reference> <candidate> [options]

flag	default	description
reference	—	reference model id/path (e.g. the fp16 model)
candidate	—	candidate model id/path (e.g. the quantized copy)
--seqs	16	eval sequences
--seqlen	256	tokens per sequence
--batch	4	sequences per forward
--device	auto	cuda or cpu
--dtype	float16	float16, bfloat16, or float32
--data-dir	data	cache dir for eval text
--out	—	write the full JSON report to this file