nx_gguf.nx

nx_gguf.nx -- GGUF v3 model-file structural parser. Reads the file format used by llama.cpp / GGML for quantized transformer models. GGUF v3 is the de facto standard for Llama / Mistral / Qwen / Phi / Gemma weights in 2024-2025. Closes the WEIGHT-LOADING gap in the substrate's inference path. With this brick, the substrate can: 1. mmap or sys_read_file the .gguf file into a byte buffer 2. nx_gguf_parse(buf, len, out_header) -- get tensor count + dims 3. Iterate nx_gguf_tensor_at(header, i) -- per-tensor metadata 4. Compose nx_qb_dequantize_tensor / nx_qb8_dequantize_tensor / nx_q4k_dequantize_tensor on the named tensors at the offsets this parser exposes ===== Format (GGUF v3) ========================================== Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md magic u32 LE = "GGUF" (0x46554747) version u32 LE = 3 tensor_count u64 LE metadata_count u64 LE metadata_kv[] (count = metadata_count): key_len u64 LE key utf8 bytes [key_len] type u32 LE (GGUF_TYPE_*) value type-dependent encoding tensor_info[] (count = tensor_count): name_len u64 LE name utf8 bytes [name_len] n_dims u32 LE dims u64 LE * n_dims type u32 LE (ggml_type: F32, F16, Q4_0, Q4_K, Q8_0, ...) offset u64 LE (byte offset into the tensor data section) -- alignment padding to GGUF_ALIGNMENT (32) bytes --