rontolisp:dequantize

(rontolisp:dequantize matrix element-type)

Expands a quantized matrix (rontolisp:quantize, a GGUF's Q8_0 tensor) into a fresh packed float array of the same dimensions and of element-type -- single-float, double-float or bfloat16. Every element is q * d, which is exact in single-float (an 8-bit quant times a binary16 scale); the bfloat16 copy narrows it to nearest even.

This is the way onto every packed-float operation a quantized matrix does not take directly: vec:matvec reads it as it is, everything else reads the dequantized copy. The copy is what the arithmetic ran over, at four bytes an element instead of one -- dequantize the tensors you need, not the checkpoint.

Interpreter and JVM only; both WASM backends refuse it at compile time.