gguf-py has been reporting the wrong byte size for Q8_1 blocks
GitHub·medium signal
llama.cpp PR #29036 (merged 2026-09-18 08:43 UTC) corrects `GGML_QUANT_SIZES[Q8_1]` in `gguf/constants.py` from `(32, 4 + 4 + 32)` to `(32, 2 + 2 + 32)`. `block_q8_1` in ggml.h stores its `d` and `s` scales as `ggml_half`, so the real struct is 36 bytes, not 40, and any Python tooling computing tensor offsets or file sizes from the constant was off by 4 bytes per block. Verified directly against `sizeof(block_q8_1)`.