| set -euo pipefail | |
| : "${LLAMA_DIFFUSION_BIN:?Set LLAMA_DIFFUSION_BIN}" | |
| : "${GGUF_PATH:?Set GGUF_PATH to a local diffusiongemma GGUF file}" | |
| PROMPT="${1:-User: Write a short haiku about GPUs.\n\nAssistant:}" | |
| MAX_TOKENS="${MAX_TOKENS:-128}" | |
| "${LLAMA_DIFFUSION_BIN}" \ | |
| -m "${GGUF_PATH}" \ | |
| -ngl "${N_GPU_LAYERS:-99}" \ | |
| -cnv \ | |
| -n "${MAX_TOKENS}" \ | |
| -p "${PROMPT}" | |