Research map / Architectures and efficiency / Quantization

QTIP: Quantization with Trellises and Incoherence Processing

NeurIPS 2024 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes QTIP, a weight-only LLM post-training quantizer that uses ultra-high-dimensional trellis-coded quantization with a hardware-efficient bitshift trellis and compute-based pseudorandom Gaussian codes for fast inference.

Problem

Vector quantization for LLM weight PTQ is limited to low dimensions (≤8) because its codebook size grows exponentially with dimension, both in quantization cost and in inference-time cache footprint.

Approach

Replace VQ with trellis-coded quantization over very long sequences (dimension >100, e.g. 256) using a bitshift trellis so that decoding is parallel and stateless, and generate the codebook values on the fly with 1-3 instruction pseudorandom Gaussian functions (1MAD, 3INST) or a small 2D LUT (HYB), so no large codebook must be stored.

Key results

• At 2 bits on Llama 2 70B, QTIP with the hybrid code reaches 3.16 W2 perplexity... (2-70 with HYB, ctx 4096: W2 3.70 on 2-7... see tables; QTIP HYB 2-bit Llama 2 7B W2 5.86 vs QuIP# 6.19 and AQLM 6.14)
• Batch-size-1 decoding on RTX 6000 Ada for Llama 2 7B at 2 bits reaches 188 tok/s, matching QuIP# (186) and above AQLM (81.5).
• On i.i.d. Gaussian at 2 bits, 256D TCQ achieves 0.068-0.071 MSE vs 0.089 for E8P VQ and 0.063 distortion-rate bound.

Measured effects reported in the paper

Surprising findings

• Computed pseudorandom codes with no fine-tuning outperform QuIP# and AQLM even with fine-tuning on almost all model sizes at 2-3 bits.
• Quantizing layer 0 v of Llama 3 70B collapses zero-shot performance without special prefix tokens.

Limitations

• Hybrid code relies on a 2KiB LUT and fine-tuning for best results; pure computed codes are not tunable.
• Throughput measured only on NVIDIA GPUs (RTX 6000 Ada, RTX 3090, A6000).
• Tail-biting approximation is heuristic; exact tail-biting is intractable for L≥12.

Experimental setup

Datasets: RedPajama WikiText2 C4

Benchmarks: WikiText2 C4 ArcC ArcE PiQA Winogrande BoolQ HellaSwag

Metrics: perplexity zero-shot accuracy decoding throughput (tokens/s) MSE on i.i.d. Gaussian

Baselines: QuIP# AQLM GPTVQ-2D PV-Tuning GPTQ AWQ OmniQuant

Builds on: QuIP# QuIP LDLQ / GPTQ-style BlockLDLQ random permutation trellis coder

Closest papers by meaning

• Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.89)
• QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · similarity 0.88)
• ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.86)
• ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.86)
• GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.86)
• PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.86)