7.22. dasllama-convert — Offline Model Prep

dasllama-convert turns a GGUF model file into a prepared .dlim image offline, so the first real load maps the image instead of converting. The image is box- and knob-specific: the default output name carries the image’s lane (model.gguf.metal.dlim, model.gguf.cpu-arm-i8mm.dlim, model.gguf.vulkan.dlim, a -mtp suffix when a split head rides the load) and is exactly what the normal dasLLAMA load path looks for; the image header carries the full bake identity, so a stale lane re-bakes in place and the lanes coexist on one disk.

7.22.1. Quick start

daslang -jit utils/dasllama-convert/main.das -- -m <model.gguf> [-o <out.dlim>] [-q q8]
  • --flavor picks which image a multi-flavor config emits (planar CPU, metal blob, vulkan-baked). -f vulkan bakes offline — no VRAM touched.

  • --dump-config prints this box’s DlimConfiguration as JSON; --config x.json bakes against a supplied one, verified fail-closed against the resulting identity — a knob that did not take is an error, never a silently different image.

Conversion is I/O bound at both ends; per-stage timings are reported so a slow box is diagnosable.

See also

dasllama-server — an OpenAI-compatible server over dasLLAMA – the inference server that loads the images