checkpoint:stage-float-bits

(checkpoint:stage-float-bits stream count format dst &key (start 0))

Reads count little-endian 16-bit words from the byte stream, at its current position, as format bit patterns -- :float16 or :bfloat16 -- and widens them into dst, a packed float array of any rank, row-major from flat index start. Returns dst.

A :bfloat16 format into a bfloat16 destination is not staged at all: those are the same two bytes, so the tensor arrives in one read-sequence with no buffer and no conversion, which is what the width is for on the load path. Every other pairing is staged. The words are staged in chunks of a million elements through one (unsigned-byte 16) buffer the package reuses for every tensor of every file, each chunk widened with rontolisp:widen-float-bits at its own :start offset. A packed integer vector costs eight bytes an element on the interpreter and the JVM, so a tensor staged whole would cost four times its file size in temporaries; taking the stream rather than a staged vector is what makes it impossible to get that wrong.

CL-USER> (with-open-file (s "model.safetensors" :element-type '(unsigned-byte 8))
           (checkpoint:skip-bytes s data-start)
           (checkpoint:stage-float-bits s (* 2048 2048) :bfloat16
                                        (checkpoint:make-tensor '(2048 2048) 'single-float)))
#<packed single-float array (2048 2048)>

Backend support

Every backend that has a filesystem: the interpreter, a compiled .class/.jar, and both WASM backends.