llama.cpp Introduces Dynamic VBR KV Cache

old-mike · reddit · 2026-07-14

This post introduces the VBR (Variable Bit Rate) KV cache, a new feature in spiritbuun's llama.cpp fork:

Instead of a fixed KV quantization level for the entire context, it sets a floor, and the cache dynamically steps down quantization levels as the context grows.
The -ct vbr command automatically calculates the available KV budget based on remaining VRAM and dynamically reduces precision during the decode process.
Tested on an RTX 3060 12GB compared to fixed quantization:
Decode TPS is only 4% slower
TTFT is roughly halved
Lower VRAM usage
Automatically adapts to a context of 216k tokens
The post also mentions a fused kernel bug, for which the author has already reverted the relevant commit.

More from Infra