Native libraries for local AI tools can conflict when several tools share one environment. The community-maintained conda-forge package installs llama.cpp binaries in a named conda environment, keeping their dependencies separate from other projects.
A working conda installation on Windows, macOS, or Linux is required. The package is named llama.cpp, while its installed command-line programs include llama-cli and llama-server.
A conda-forge-only solve keeps the compiler runtime and supporting libraries in the same channel. A separate named environment also keeps these dependencies out of the base conda environment.
Steps to install llama.cpp in a conda environment:
- Install llama.cpp in a new environment named llama-cpp from conda-forge.
$ conda create --yes --name llama-cpp --override-channels --channel conda-forge llama.cpp Channels: - conda-forge Platform: linux-aarch64 ##### snipped ##### environment location: /opt/conda/envs/llama-cpp ##### snipped ##### llama.cpp conda-forge/linux-aarch64::llama.cpp-11860-cpu_openblas_h10e8b54_0 ##### snipped ##### Executing transaction: ...working... done
The llama.cpp value is the package name; llama-cpp is only the local environment name. The selected version and build vary by platform and available packages. A new environment name avoids changing packages in an existing project, and --override-channels applies only to this solve without changing saved channel settings.
- Confirm that conda recorded the package in the new environment.
$ conda list --name llama-cpp llama.cpp # packages in environment at /opt/conda/envs/llama-cpp: # # Name Version Build Channel llama.cpp 11860 cpu_openblas_h10e8b54_0 conda-forge
- Verify that llama-cli starts from the environment.
$ conda run --name llama-cpp llama-cli --version version: 0.7.0-dev (build 1, commit a553e84) built with GNU 15.3.0 for Linux aarch64
- Verify that llama-server starts from the same environment.
$ conda run --name llama-cpp llama-server --version version: 0.7.0-dev (build 1, commit a553e84) built with GNU 15.3.0 for Linux aarch64
Both commands should return a version and compiler line without a missing-library error. The package version and binary build number can differ. conda run --name llama-cpp selects the installed environment without activating it in the current shell. The package contains the binaries; model weights come from a separate GGUF file.
Related: Run a local GGUF model with llama.cpp