Research map / Architectures and efficiency / Quantization
Channel-Aware Mixed-Precision Quantization for Efficient Long-Context Inference
Chengxi Liao, Zeyi Wen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieval channel 4-bit allocation to 2-bit baseline (ablation) | ChanMix2 · RULER (average) Llama-3.1 | accuracy: 75.40 -> 81.50 | Improved |
| add retrieval channel allocation to 2-bit baseline (ablation) | ChanMix2 · RULER (average) Mistral | accuracy: 72.13 -> 86.12 | Improved |
| add outlier channel 3-bit compensation to 2-bit baseline (ablation) | ChanMix2 · RULER (average) Llama-3.1 | accuracy: 75.40 -> 81.13 | Improved |
| add outlier compensation to 2-bit baseline (ablation) | ChanMix2 · RULER (average) Mistral | accuracy: 72.13 -> 84.17 | Improved |
| add outlier channels on top of retrieval channels (ablation) | ChanMix2+R · RULER (average) Llama-3.1 | accuracy: 81.50 -> 82.16 | Improved |
| add outlier channels on top of retrieval channels (ablation) | ChanMix2+R · RULER (average) Mistral | accuracy: 86.12 -> 86.35 | Improved |
| bit reallocation versus uniform 2-bit at group 128, residual 32 (ablation) | ChanMix · RULER Llama-3.1 | accuracy: 69.78 -> 78.02 | Improved |
| larger group size 128 instead of 32 at residual 32 (ablation) | ChanMix · RULER Llama-3.1 | accuracy: 83.92 -> 78.02 | Hurt |
| vary residual length 32/64/128 at group 32 (ablation) | ChanMix · RULER Llama-3.1 | accuracy: 83.92, 84, 83.87 | No clear effect |
| channel-aware mixed precision versus KIVI | ChanMix · RULER Llama-3.1 | accuracy: 76.51 -> 82.16 | Improved |
| ChanMix versus KIVI on short-context GSM8K | ChanMix · GSM8K Llama-3.1 | accuracy: 74.75 -> 73.54 | Hurt |
| ChanMix versus KIVI on reasoning AIME24 | ChanMix · AIME24 DS-R1-Llama | pass@1: 26.67 -> 26.67 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ZipCache — ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · similarity 0.93)
- KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference (ICML 2025 · similarity 0.91)
- Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models (ICML 2025 · similarity 0.90)
- NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache (NeurIPS 2025 · similarity 0.90)
- Progressive Mixed-Precision Decoding for Efficient LLM Inference (ICLR 2025 · similarity 0.90)
- KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache (ICML 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
