Usage

One-Time Setup

./scripts/bootstrap_macos_llamacpp.sh
./scripts/download_diffusiongemma_q4.sh

The downloaded Q4_K_M model is about 16 GB and is intentionally ignored by git.

./scripts/ask_matrix_gemma.sh "What is the weirdest fact about black holes?"

MATRIX_BIAS=8 MATRIX_FRACTION=0.25 MATRIX_STEPS=48 \
  ./scripts/ask_matrix_gemma.sh "Explain discrete diffusion in one paragraph."

Environment variables:

MATRIX_BIAS: additive kana-token logit bias at the start. Try 6 to 14.
MATRIX_FRACTION: portion of denoising schedule affected by Matrix conditioning. Try 0.20 to 0.35.
MATRIX_STEPS: max entropy-bound denoising steps. Use 48 for quality.
MATRIX_BLOCKS: number of 256-token canvases to generate. Use 1 for short answers.
MATRIX_VISUAL_INTERVAL: redraw every Nth step. Use 1 for full animation.
N_GPU_LAYERS: llama.cpp GPU layer offload. Default is 99.
N_PREDICT: requested token budget. Default is 256.
DIFFUSIONGEMMA_MODEL: override the GGUF path.

MATRIX_STEPS=12 MATRIX_BIAS=14 MATRIX_FRACTION=0.45 \
  ./scripts/ask_matrix_gemma.sh "What is 2+2? Answer briefly."

Low step counts make the Matrix effect easy to see but reduce answer quality.