[Afloat16]

What & why

Numeric GGUF arrays performed one Python read/dispatch per scalar, making vocabulary-sized score/type tables expensive.

Decode numeric arrays in blocks of at most 8192 elements (64 KiB), preserving scalar types, little-endian decoding, empty arrays and truncation errors. String arrays retain their path.

Reference: https://github.com/ggml-org/ggml/blob/cd6d2b75f5d3e4b2f3fe8e8fbc6c8b84ac80e74a/docs/gguf.md

Area

Model asset tools (edge0-infer/tools/assets/).

How was it tested?

python -m pytest -q tools/assets/tests/test_gguf_lite.py

16 tests passed with Python 3.12 and NumPy 2.3.5. A 25000-element uint32 table uses 13 reader calls instead of 25009, with a maximum read of 32768 bytes. The baseline fails the read-count bound. All 11 scalar formats, 8193-element/empty arrays, truncated payloads, existing tensor round-trips and quantization are covered.

All 14 L0 checks passed in a local integration with the other focused fixes. git diff --check passed.

Checklist

• Full ctest --preset cpu -L "L0|L1" passes locally
• Added/updated focused regressions and ran the affected target
• No public interface/configuration change requiring additional documentation