@JohannesGaessler applied your suggested changes. I've moved the GGML_ASSERT just after the endianness check, otherwise it will trigger the assert before the check could produce any useful information for the end-user.

PTAL again.

P.S., GGML_ASSERT fails because the version field in non-native endian models are read as 50331648.

Edit: Also re-tested the new changes on ARM64 and s390x. Both worked as intended.

Unfortunately that GGML_ASSERT is still required because when I change the version to 0x0001 FFFF, the assert catches it via

llama.cpp-s390x/ggml/src/gguf.cpp:363: GGML_ASSERT(ctx->version > 0 && ctx->version <= 65535) failed

while the endianness check did not; as expected.

If the CI passes, feel free to merge it into master :)