EigenKV: Extendiendo Ventanas de Contexto con Compresión Inteligente de KV-Cache

The KV-Cache Bottleneck

Modern large language models use a key-value cache to store intermediate computations during text generation. This cache grows linearly with context length, creating a fundamental bottleneck for long-context applications.

For a 70B parameter model processing 128K tokens, the KV-cache alone can consume over 32GB of GPU memory—often more than the model weights themselves.

Understanding KV-Cache Growth

The memory consumption of KV-cache follows a predictable pattern: each new token adds a fixed amount of memory per layer. For long-context applications, this quickly becomes the dominant memory consumer.

This growth pattern forces uncomfortable tradeoffs: either limit context length, upgrade to more expensive hardware, or sacrifice batch size.

Our Solution

EigenKV applies structured compression to the KV-cache during generation. Unlike methods that simply evict old tokens, we preserve information from the entire context while reducing memory footprint.

Key Features

Training-Free

EigenKV works with any transformer-based model out of the box. No fine-tuning, no modified architectures—just plug in and benefit from reduced memory usage.

Quality Preservation

Our compression is designed to preserve the information most relevant to generation quality. Benchmark results show minimal degradation across diverse tasks.

Streaming-Compatible

EigenKV operates in streaming mode, compressing cache entries as they are created. This means no sudden memory spikes or batch processing requirements.

Use Cases

EigenKV enables several previously difficult scenarios: Document QA over 100K+ token documents on consumer GPUs, multi-turn conversations with full context retention, code generation with large repository context, and reducing inference costs for long-context applications.