error: could not find native static library `culibos`

[xutianyi1999]

Windows should not need to link culibos, and the lib is not found in the CUDA path.

llama-cpp-rs/llama-cpp-sys-2/build.rs

Line 420 in 6ab83db

println!("cargo:rustc-link-lib=static=culibos");

[alexivai]

go to llama-cpp-sys-2 and open build.rs
Replace to this

if cfg!(feature = "cuda") && !build_shared_libs {
    println!("cargo:rerun-if-env-changed=CUDA_PATH");
    for lib_dir in find_cuda_helper::find_cuda_lib_dirs() {
        println!("cargo:rustc-link-search=native={}", lib_dir.display());
    }
    println!("cargo:rustc-link-lib=cudart"); // Только динамическая
    if matches!(target_os, TargetOs::Windows(_)) {
        println!("cargo:rustc-link-lib=cublas");
        println!("cargo:rustc-link-lib=cublasLt");
    } else {
        println!("cargo:rustc-link-lib=cublas");
        println!("cargo:rustc-link-lib=cublasLt");
    }
    if !cfg!(feature = "cuda-no-vmm") {
        println!("cargo:rustc-link-lib=cuda");
    }
}

[DAIMENGEN]

This build script is modified to properly handle platform-specific CUDA linking:

On Windows:

NVIDIA’s CUDA SDK does not ship culibos.lib, and static libraries like cublas_static.lib are typically unavailable.

Attempting to statically link cudart_static on Windows will fail due to missing symbols provided by culibos.

Therefore, on Windows, dynamic linking is preferred. The .lib files (like cudart.lib) act as import libraries for corresponding .dll files (e.g., cudart64_*.dll).

if cfg!(feature = "cuda") && !build_shared_libs {
    // Re-run build script if CUDA_PATH environment variable changes
    println!("cargo:rerun-if-env-changed=CUDA_PATH");

    // Add CUDA library directories to the linker search path
    for lib_dir in find_cuda_helper::find_cuda_lib_dirs() {
        println!("cargo:rustc-link-search=native={}", lib_dir.display());
    }

    // Platform-specific linking
    if cfg!(target_os = "windows") {
        // ✅ On Windows, use dynamic linking.
        // Static linking is problematic because NVIDIA does not provide culibos.lib,
        // and static CUDA libraries (like cublas_static.lib) are usually not shipped.

        println!("cargo:rustc-link-lib=cudart");       // Links to cudart64_*.dll
        println!("cargo:rustc-link-lib=cublas");       // Links to cublas64_*.dll
        println!("cargo:rustc-link-lib=cublasLt");     // Links to cublasLt64_*.dll

        // Link to CUDA driver API (nvcuda.dll via cuda.lib)
        if !cfg!(feature = "cuda-no-vmm") {
            println!("cargo:rustc-link-lib=cuda");
        }
    } else {
        // ✅ On non-Windows platforms (e.g., Linux), static linking is preferred and supported.
        // Static libraries like cudart_static and cublas_static depend on culibos.

        println!("cargo:rustc-link-lib=static=cudart_static");
        println!("cargo:rustc-link-lib=static=cublas_static");
        println!("cargo:rustc-link-lib=static=cublasLt_static");

        // Link to CUDA driver API (libcuda.so)
        if !cfg!(feature = "cuda-no-vmm") {
            println!("cargo:rustc-link-lib=cuda");
        }

        // culibos is required when statically linking cudart_static
        println!("cargo:rustc-link-lib=static=culibos");
    }
}

[DAIMENGEN]

Looks like this issue has been fixed in commit #759. It should be safe to close this now.
@xutianyi1999 @MarcusDunn

[MarcusDunn]

Thanks for the ping