Misc. bug: llama-cli's inference result seems not correct on 64bit-Windows

[zhouwg]

Name and Version

$ ./llama-cli --version
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
ggml_backend_load_best: search path does not exist
version: 4894 (8ae58b3)
built with cc (GCC) 15.0.1 20250223 (experimental) for x86_64-pc-cygwin

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

llama-cli

Command line

./llama-cli -m /cygdrive/z/zhouwg/gemma-2b.Q8_0.gguf -p "pls tell me 1+1=?"

Problem description & steps to reproduce

recently I tried a new approach of build llama.cpp with cygwin on Windows 64 without MS's huge IDE, details could be found at a draft PR:#12215, it seems the inference result is not correct as expected:

llama-cli-result-not-correct-on-cygwin-x86-64-windows.mp4

I don't understand the reason and help from community is greatly appreciated, thanks.

First Bad Commit

No response

Relevant log output

[zhouwg]

fixed, details could be fount at #12215