Skip to content

TurboPress

Certified low-bit LLM weight quantization. TurboPress compresses the decoder linears of a Hugging Face causal LM (Llama / Qwen / Mistral family) to 2–6 bits per weight with a measured, reproducible pipeline:

  1. Seeded randomized rotation (block fast Walsh–Hadamard × stored signs) to decorrelate weight residuals from activation structure.
  2. Rotation-aware activation equilibration — s_j = m_j^α / c_j^(1/2−α) with the derived optimal exponent: α = ¼ for a feedback-free quantizer (Proposition 1; not the AWQ square-root rule) and α = ⅛ with error feedback on (Proposition 2 — feedback absorbs activation structure).
  3. Trellis-coded quantization with a data-free, trellis-optimized codebook (Viterbi search over a rate-1/2 convolutional code).
  4. Block-LDLQ error feedback over the trellis (v0.5.0, on by default) — each column block's Hessian-weighted error feeds forward while the packed format stays identical. On SmolLM2-135M at 3 bits this cut KL 14.6% and perplexity 8.1% vs v0.4.2 at the same artifact size.

Every run validates against the full-precision model and reports KL, top-1 next-token agreement, and perplexity, then writes a self-contained artifact that reloads without the repo.

pip install "turbopress[llm]"
turbopress compress Qwen/Qwen3-4B --bits 3
turbopress validate Qwen/Qwen3-4B ./turbopress_out/Qwen3-4B-turbopress-3bit

See the Quickstart to get running, the CLI reference for every flag, and Research for what was measured — including what was refuted (unbiased QJL residual correction of weights) as well as what worked.

Research prototype

TurboPress is an actively developed research prototype. Read the Limitations before relying on it in production.