nxgguf.nx

nxgguf.nx -- Nishi sovereign tensor weight format. Replaces HuggingFace safetensors / GGUF with content-addressed, memory-mapped, dedupable weight blocks. Single-binary deploy of any model; instant cold-load via mmap; LoRAs share base weights at the storage layer. Format overview: [magic 4] = "NXGF" [version 4] = u32 LE, currently 1 [tensor_count 8] = u64 LE, total tensors [metadata_offset 8] = u64 LE, offset to metadata block [metadata_length 8] = u64 LE, byte length of metadata [block_table_offset 8] = u64 LE, offset to block dir [block_table_length 8] = u64 LE, byte length of block dir [data_offset 8] = u64 LE, where weight bytes start [reserved 16] = zeros = 64 bytes total header metadata block: protobuf-encoded TensorRegistry repeated TensorEntry { string name = 1; // "model.layers.0.attn.q.weight" repeated int64 shape = 2; // [512, 512] int32 dtype = 3; // 0=fp32 1=fp16 2=bf16 3=fp8 4=int8 5=int4 string block_hash = 4; // sha256 of the tensor's bytes int64 byte_length = 5; // raw byte count } block table: array of BlockDescriptor bytes hash (32) // sha256 uint64 file_offset ( 8) // where in file uint64 byte_length ( 8) uint32 ref_count ( 4) // # tensors pointing here uint32 reserved ( 4) = 56 bytes per block descriptor data section: concatenated raw weight bytes; tensor reads `block.byte_length` bytes starting at `block.file_offset`.