BeeLlama Main

beellama/main

beellama/main adds this fork's memory controls to BeeLlama v0.4.3 (ba27edad2a). It does not descend from llama/main. Documented checkpoint: 4b86269fdf.

Note

BACKEND SUPPORT VARIES: Fork placement controls require a compatible backend. BeeLlama features such as KVarN keep their own build and backend requirements.

BeeLlama features

Use BeeLlama's v0.4.3 feature guide and v0.4.3 argument reference. Newer BeeLlama documentation may list options absent from this pinned base.

Fork additions

• Pinned host KV
• Partial target and draft KV residency
• Independent recurrent-state placement
• Phase-aware and live-context workspace sizing
• Independent draft ubatch
• Capped MTP recurrent planes and sparse snapshots
• Canonical host-KV quantization
• Validation tools, CUDA VMM telemetry, and perplexity capacity correction

These controls trade memory against transfers, workspace growth, or MTP replay. Benchmark the exact BeeLlama model and backend after enabling them.

Build

git clone --branch beellama/main https://github.com/GenerelSchwerz/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j 20
./build/bin/llama-server --help

Add KVarN or other BeeLlama build switches from the version-matched BeeLlama guide.

Fork flags and setup - Hardware guides - Feature index - Home