tests: disable GGUF test for bad value size

[JohannesGaessler]

See #10830 (comment) .

This PR makes the test failure with a bad value size deterministic (as in independent of the seed). The way this test causes a failure is not really well-defined since it just writes a random wrong number of wrong, random bytes to the file. So depending on the RNG you can get a random, uncaught failure in a calloc call. This PR makes the test failure deterministic to prevent false positives in the CI until I have the C++ refactor ready.

@ggerganov where in llama.cpp is the sanitizer logic defined? I usually use Valgrind for my local development but I'd like to assert that there aren't any more random failures with the method used in the CI.

[ggerganov]

You can enable sanitizers like this:

# address + undefined sanitizers enabled
cmake [...] -DLLAMA_SANITIZE_ADDRESS=ON -DLLAMA_SANITIZE_UNDEFINED=ON

# thread sanitizer enabled (this one is not compatible with the address sanitizer)
cmake [...] -DLLAMA_SANITIZE_THREAD=ON

[JohannesGaessler]

It seems that this problem is caused by the allocation attempt exceeding the maximum allowed allocation size of the sanitizer (1 TiB) but with a size which is still below the threshold of what could possibly be allocated given the address space (this case can be detected prior to allocation). When running the code in production what should happen is that the allocation will fail due to hardware constraints, this is then detected as an error and NULL is returned. So I think for now we should just disable this test and I'll look into how to fix it for the C++ refactor.

[ymcki]

Thanks for the fix. I too got the mysterious calloc crash during test-gguf