Fuse the adapter at adapterDir into the base model at modelDir and
write a complete, loadable snapshot to outDir (weights + config.json +
tokenizer aux files). Quantization layout is preserved verbatim: fused
modules are dequantized, updated, and re-quantized with their own source
spec; untouched modules pass through bit-identical. With
options.dequantize every quantized module is written dense instead.
OptionalonProgress: (e: FuseProgress) => void
Resolves a query the way the catalog does (a model directory as given, else a downloaded model); the write token and the push behind
--upload-repoare the catalog's too.