safetensors:read

(safetensors:read path &key only element-type)

Reads a safetensors checkpoint into a hash table (equal, string keys) of tensor name -> packed float array of the file's shape: a rank-1 vector for a one-dimensional tensor, a rank-N packed array otherwise. path is a .safetensors file, a model.safetensors.index.json (its shards are each opened once and walked once), or a directory holding either. element-type is single-float (the default), double-float, or bfloat16 on the interpreter and the JVM (every other backend refuses that width by name); F32 tensors are read straight in, F16 and BF16 through the checkpoint staging, and any other dtype signals an error naming the tensor and the dtype. A BF16 tensor into a bfloat16 destination is the one pairing with no conversion at all -- the file's own two bytes an element, in one transfer. only is a predicate over the name: a tensor it rejects is skipped in bounded reads rather than staged, which is how a multimodal checkpoint's tower and speculative head are left on disk.

CL-USER> (defparameter *w* (safetensors:read "TinyLlama-1.1B-Chat-v1.0"
                                             :only (lambda (name) (search "layers.0." name))))
*W*
CL-USER> (hash-table-count *w*)
9
CL-USER> (array-dimensions (gethash "model.layers.0.self_attn.q_proj.weight" *w*))
(2048 2048)
CL-USER> (array-element-type (gethash "model.layers.0.input_layernorm.weight" *w*))
SINGLE-FLOAT

Backend support

Every backend that has a filesystem: the interpreter, a compiled .class/.jar, and both WASM backends. What is narrower is the bfloat16 destination element type, which stays interpreter and JVM only (every other backend refuses that width by name).