Research map / Vision-language and multimodal / Multimodal LLMs
Efficient Multimodal Large Language Model via Dynamic KV Cache Quantization
Jiahao Fan, Chien-Ming Chen
Read the paperCheck your idea against 137,217 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 a dynamic KV cache quantization approach to reduce the memory and compute cost of multimodal LLM inference for resource-constrained deployment.
Problem
Multimodal LLMs face high computational cost and KV-cache memory footprint during inference, hindering real-time deployment on resource-constrained devices.
Approach
Applies dynamic quantization to the KV cache of multimodal LLMs to shrink its memory footprint during inference; specific mechanism not stated in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add fine-grained grouping (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA | accuracy (%): 57.64 -> 58.01 | Improved |
| add fine-grained grouping (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · GQA | accuracy (%): 61.47 -> 62.13 | Improved |
| add token/channel recorders and scaling (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · GQA | accuracy (%): 61.75 -> 62.13 | Improved |
| add token/channel recorders and scaling (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA | accuracy (%): 57.82 -> 58.01 | Improved |
| decrease group size from 128 to 32 (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA | accuracy (%): 57.43 -> 58.01 | Improved |
| decrease group size from 32 to 16 (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA | accuracy (%): 58.01 -> 58.05 | No clear effect |
| use per-token instead of per-channel grouping for K (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · GQA | accuracy (%): 61.92 -> 62.05 | Mixed |
| use per-token instead of per-channel grouping for V (ablation) | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · GQA | accuracy (%): 61.84 -> 62.13 | Improved |
| use Dynamic KV Cache Quantization instead of uniform 4-bit KV quantization | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA (LLaVA-1.5-7B) | accuracy (%): 0.12 -> 58.01 | Improved |
| use Dynamic KV Cache Quantization instead of SKVQ | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · SciQA (LLaVA-1.5-7B) | accuracy (%): 56.02 -> 69.74 | Improved |
| use Dynamic KV Cache Quantization instead of KIVI | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA (Qwen-VL) | accuracy (%): 63.87 -> 63.92 | No clear effect |
| use Dynamic KV Cache Quantization instead of MiKV | Dynamic KV Cache Quantization (LLaVA-1.5-7B) · TextVQA (LLaVA-1.5-13B) | accuracy (%): 60.87 -> 61.12 | 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.
- ZIPVL — ZipVL: Accelerating Vision-Language Models through Dynamic Token Sparsity (ICCV 2025) · better on 2 of 2 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)
- Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models (ICML 2025 · similarity 0.92)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.91)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.91)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.91)
- Q Cache: Visual Attention Is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model (AAAI 2026 · similarity 0.91)
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.
