How to fix: llama.cpp built without libcurl

[eonward]

Hi!

It seems like my llama.cpp can't use libcurl in my system.

When I try to pull a model from HF, I get the following:
llama_load_model_from_hf: llama.cpp built without libcurl, downloading from Hugging Face not supported.

I'm on Ubuntu, and have the following modules installed:
libcurl3t64-gnutls
libcurl4t64

libcurl4t64 in particular provides libcurl4

➜  llama.cpp git:(master) dpkg -L libcurl4t64
/usr/lib/x86_64-linux-gnu/libcurl.so.4.8.0
/usr/share/doc/libcurl4t64/changelog.Debian.gz
/usr/share/lintian/overrides/libcurl4t64
/usr/lib/x86_64-linux-gnu/libcurl.so.4

(abridged)

Is this the curl used in common_download_file in common.cpp, or is that looking for something else?

Any ideas? Thanks!

[danbev]

Can you double check that the llama.cpp version that you build used the LLAMA_CURL flag?

If using cmake this would look something like this:

$ cmake -S . -B build -DLLAMA_CURL=ON

Or if using make:

$ make LLAMA_CURL=1

Also, if I remember correctly you might need the libcurl4-openssl-dev package as well.

$ sudo apt install libcurl4-openssl-dev

[brauliobo]

worked here

[cryptonight-creator]

Worked for me too 😊😊

[cyberluke]

Why this is not in BUILD.md ? And this seems to be required, not optional. Why this is not in Makefile? (latest github version 18.2.2025)

[cyberluke]

why you presume fetching models is not default feature?

[cyberluke]

The question was not only about Makefile, but also CMakefile. The issue is generic, not specific to Make or CMake.

[cyberluke]

I point your knowledge of build etiquette is completely wrong. LLAMA_CURL is stupid solution....but then again what one can expect from LLAMA, right? In that case best effort, great work!

[danbev]

why you presume fetching models is not default feature?

I think this has to do with CuRL and that it might not be available on all systems that are supported.

[cyberluke]

1. nowadays CuRL is even on Windows
2. if that -> implement fail-over-solution for CuRL not avail

[tilkinsc]

Since this discussion didn't have any solutions one might be looking for on windows, I installed vcpkg

git clone https://github.com/microsoft/vcpkg.git
cd vcpkg
bootstrap-vcpkg.bat

Then I installed libcurl

vcpkg install curl:x64-windows

Then I integrated vcpkg so that libs can be automatically found

vcpkg integrate install

From there it emitted a cmake file I could pass to cmake via a -D

-DCMAKE_TOOLCHAIN_FILE=C:/code/vcpkg/scripts/buildsystems/vcpkg.cmake

And then I built llama.cpp

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build -DLLAMA_CURL=ON -DGGML_CUDA=ON -DCMAKE_TOOLCHAIN_FILE=C:/code/vcpkg/scripts/buildsystems/vcpkg.cmake
cmake --build build --config Release

You might get away with using ninja, but I don't think it really matters since most of the time compiling is spent in .cu (cuda) files.