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]
--flavorpicks which image a multi-flavor config emits (planar CPU, metal blob, vulkan-baked).-f vulkanbakes offline — no VRAM touched.--dump-configprints this box’sDlimConfigurationas JSON;--config x.jsonbakes 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