Module pack 

Pack qwen4-exp checkpoints into aligned files. MLX Q4 tensors are copied bit for bit; native BF16 tensors are converted as the files are written.

dense.bin every tensor that is not an expert matrix, an n-gram shard, or vision (64-byte aligned, name order) experts.bin [layer][expert] records (see ExpertLayout) ngram.bin [row] records of weight | scales | biases manifest.json

Modules

affine 🔒

Affine Q4 groups: eight low-first nibbles per word, BF16 scale and bias.

source 🔒

Lazy tensor views let both checkpoint formats use the same packed writer. BF16 conversion holds one quantization group at a time, including when a fused expert tensor spans an entire layer. No converted checkpoint is staged.

Structs

Progress 🔒

Enums

Class 🔒

Constants

CHAT_METADATA_FILES 🔒

Auxiliary files copied without replacing metadata already present in a store.

DENSE_ALIGN 🔒

Functions

classify 🔒

copy_chat_metadata 🔒

Older packed directories can acquire template metadata without repacking weights.

dtype_name 🔒

missing_chat_metadata 🔒

Whether an available source has metadata that the prepared store lacks.

pack

pack_dense 🔒

pack_experts 🔒

pack_ngram 🔒

pad_to 🔒

prepare

Prepare selected precisions without loading the inference engine. Low-bit stores always derive from the aligned Q4 base, which is retained for reuse.