Issues / #3043

#3043 [Bug]: strata-gguf --check returns success for invalid architecture and wrapped tensor bounds

open · @j-luwierski · 0 comentários · No GitHub

NVIDIA / CUDAModels & quantsLinux

Descrição

### What happened

On upstream main `477ac13c`, `strata-gguf FILE --check` returns exit status 0 for failed checks:

- With `general.architecture=llama`, it prints `architecture guard: FAIL` but returns 0. Missing architecture metadata has the same result.
- For a single F32[1] tensor in a four-byte payload, offset `UINT64_MAX - 3` wraps `payload - offset` to 8. The geometry bracket reports zero failures and returns 0.
- Changing that tensor dimension to `2^62` wraps the F32 byte count to zero and also passes the bracket.

Expected: status 1 for failed architecture checks, out-of-file known tensor payloads and overflowing tensor byte counts.

A regression invoking the real C++ CLI entry point on synthetic GGUFs fails in these four cases on upstream and passes with the proposed fix. It also covers valid files, ordinary truncation and inspection without `--check`.

### Strata version, GPU, OS

Upstream main `477ac13c` (0.1.45), Linux x86_64, GCC 14.2.0. CPU-only reproduction; no model weights or GPU required.

### Engine log

```text
unsupported architecture: FAIL (exit 0, expected 1)
missing architecture: FAIL (exit 0, expected 1)
offset that wraps the last gap to eight bytes: FAIL (exit 0, expected 1)
tensor byte count that wraps to zero: FAIL (exit 0, expected 1)
```

Searched open issues and PRs for strata-gguf, gguf_reader.cpp, architecture checks and offsets. Related #2288 covers the Python reader and #2475 covers the Metal reader; this report concerns only src/artifact/gguf_reader.cpp CLI validation.

No site

Links install, modelos, releases.