Spiritbuun's llama.cpp fork adds VBR KV cache for dynamic context

Spiritbuun has added a Variable Bit Rate (VBR) feature to their llama.cpp fork, allowing the KV cache quantization tier to degrade dynamically as context length increases. This enables users to set a minimum precision floor while automatically adjusting tiers to fit within available VRAM.

• VBR derives a KV VRAM budget from remaining memory at startup and degrades tiers on the fly as context fills.
• Benchmarks on an RTX 3060 with Qwen3.6-35B-A3B show VBR achieves nearly 2x faster TTFT (93ms vs 178ms) and lower VRAM usage (10.2 GiB vs 11 GiB) compared to fixed turbo4.
• Decode TPS is only ~4% slower with VBR (50.85 vs 52.83), and the context budget auto-fitted to 216k tokens.

This feature helps users push local inference on limited VRAM by providing dynamic context support without manual configuration of codec or context length.