llamafile lets you distribute and run LLMs with a single file (blog post)

Our goal is to make the "build once anywhere, run anywhere" dream come true for AI developers. We're doing that by combining llama.cpp with Cosmopolitan Libc into one framework that lets you build apps for LLMs as a single-file artifact that runs locally on most PCs and servers.

First, your llamafiles can run on multiple CPU microarchitectures. We added runtime dispatching to llama.cpp that lets new Intel systems use modern CPU features without trading away support for older computers.

Secondly, your llamafiles can run on multiple CPU architectures. We do that by concatenating AMD64 and ARM64 builds with a shell script that launches the appropriate one. Our file format is compatible with WIN32 and most UNIX shells. It's also able to be easily converted (by either you or your users) to the platform-native format, whenever required.

Thirdly, your llamafiles can run on six OSes (macOS, Windows, Linux, FreeBSD, OpenBSD, and NetBSD). You'll only need to build your code once, using a Linux-style toolchain. The GCC-based compiler we provide is itself an Actually Portable Executable, so you can build your software for all six OSes from the comfort of whichever one you prefer most for development.

Lastly, the weights for your LLM can be embedded within your llamafile. We added support for PKZIP to the GGML library. This lets uncompressed weights be mapped directly into memory, similar to a self-extracting archive. It enables quantized weights distributed online to be prefixed with a compatible version of the llama.cpp software, thereby ensuring its originally observed behaviors can be reproduced indefinitely.

We provide example binaries that embed several different models. You can download these from Hugging Face via the links below. "Command-line binaries" run from the command line, just as if you were invoking llama.cpp's "main" function manually. "Server binaries" launch a local web server (at 127.0.0.1:8080) that provides a web-based chatbot.

You can also also download just the llamafile software (without any weights included) from our releases page, or directly in your terminal or command prompt. This is mandatory currently on Windows.

curl -L https://github.com/Mozilla-Ocho/llamafile/releases/download/0.1/llamafile-server-0.1 >llamafile
chmod +x llamafile
./llamafile --help
./llamafile -m ~/weights/foo.gguf

On macOS with Apple Silicon you need to have Xcode installed for llamafile to be able to bootstrap itself.

On Windows, you may need to rename llamafile to llamafile.exe in order for it to run. Windows also has a maximum file size limit of 2GB for executables. You need to have llamafile and your weights be separate files on the Windows platform.

If you use zsh and have trouble running llamafile, try saying sh -c ./llamafile. This is due to a bug that was fixed in zsh 5.9+. The same is the case for Python subprocess, old versions of Fish, etc.

On Linux binfmt_misc has been known to cause problems. You can fix that by installing the actually portable executable interpreter.

sudo wget -O /usr/bin/ape https://cosmo.zip/pub/cosmos/bin/ape-$(uname -m).elf
sudo sh -c "echo ':APE:M::MZqFpD::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"
sudo sh -c "echo ':APE-jart:M::jartsr::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"

On Apple Silicon, everything should just work if Xcode is installed.

On Linux, Nvidia cuBLAS GPU support will be compiled on the fly if (1) you have the cc compiler installed, (2) you pass the --n-gpu-layers 35 flag (or whatever value is appropriate) to enable GPU, and (3) the CUDA developer toolkit is installed on your machine and the nvcc compiler is on your path.

On Windows, that usually means you need to open up the MSVC x64 native command prompt and run llamafile there, for the first invocation, so it can build a DLL with native GPU support. After that, $CUDA_PATH/bin still usually needs to be on the $PATH so the GGML DLL can find its other CUDA dependencies.

In the event that GPU support couldn't be compiled and dynamically linked on the fly for any reason, llamafile will fall back to CPU inference.

Here's how to build llamafile from source. First, you need the cosmocc toolchain, which is a fat portable binary version of GCC. Here's how you can download the latest release and add it to your path.

mkdir -p cosmocc
cd cosmocc
curl -L https://github.com/jart/cosmopolitan/releases/download/3.1.1/cosmocc-3.1.1.zip >cosmocc.zip
unzip cosmocc.zip
cd ..
export PATH="$PWD/cosmocc/bin:$PATH"

You can now build the llamafile repository by running make:

make -j8

Here's an example of how to generate code for a libc function using the llama.cpp command line interface, utilizing WizardCoder-Python-13B (license: LLaMA 2) weights.

make -j8 o//llama.cpp/main/main
o//llama.cpp/main/main \
  -m ~/weights/wizardcoder-python-13b-v1.0.Q8_0.gguf \
  --temp 0 \
  -r $'```\n' \
  -p $'```c\nvoid *memcpy_sse2(char *dst, const char *src, size_t size) {\n'

Here's a similar example that instead utilizes Mistral-7B-Instruct (license: Apache 2.0) weights.

make -j8 o//llama.cpp/main/main
o//llama.cpp/main/main \
  -m ~/weights/mistral-7b-instruct-v0.1.Q4_K_M.gguf \
  --temp 0.7 \
  -r $'\n' \
  -p $'### Instruction: Write a story about llamas\n### Response:\n'

Here's an example of how to run llama.cpp's built-in HTTP server in such a way that the weights are embedded inside the executable. This example uses LLaVA v1.5-7B (license: LLaMA, OpenAI), a multimodal LLM that works with llama.cpp's recently-added support for image inputs.

make -j8

o//llamafile/zipalign -j0 \
  o//llama.cpp/server/server \
  ~/weights/llava-v1.5-7b-Q8_0.gguf \
  ~/weights/llava-v1.5-7b-mmproj-Q8_0.gguf

o//llama.cpp/server/server \
  -m llava-v1.5-7b-Q8_0.gguf \
  --mmproj llava-v1.5-7b-mmproj-Q8_0.gguf \
  --host 0.0.0.0

The above command will launch a browser tab on your personal computer to display a web interface. It lets you chat with your LLM and upload images to it.

If you want to be able to just say:

./server

...and have it run the web server without having to specify arguments (for the paths you already know are in there), then you can add a special .args to the zip archive, which specifies the default arguments. In this case, we're going to try our luck with the normal zip command, which requires we temporarily rename the file. First, let's create the arguments file:

cat <<EOF >.args
-m
llava-v1.5-7b-Q8_0.gguf
--mmproj
llava-v1.5-7b-mmproj-Q8_0.gguf
--host
0.0.0.0
...
EOF

As we can see above, there's one argument per line. The ... argument optionally specifies where any additional CLI arguments passed by the user are to be inserted. Next, we'll add the argument file to the executable:

mv o//llama.cpp/server/server server.com
zip server.com .args
mv server.com server
./server

Congratulations. You've just made your own LLM executable that's easy to share with your friends.

(Note that the examples provided above are not endorsements or recommendations of specific models, licenses, or data sets on the part of Mozilla.)

SYNOPSIS

  o//llamafile/zipalign ZIP FILE...

DESCRIPTION

  Adds aligned uncompressed files to PKZIP archive

  This tool is designed to concatenate gigabytes of LLM weights to an
  executable. This command goes 10x faster than `zip -j0`. Unlike zip
  you are not required to use the .com file extension for it to work.
  But most importantly, this tool has a flag that lets you insert zip
  files that are aligned on a specific boundary. The result is things
  like GPUs that have specific memory alignment requirements will now
  be able to perform math directly on the zip file's mmap()'d weights

FLAGS

  -h        help
  -N        nondeterministic mode
  -a INT    alignment (default 65536)
  -j        strip directory components
  -0        store uncompressed (currently default)

Here is a succinct overview of the tricks we used to create the fattest executable format ever. The long story short is llamafile is a shell script that launches itself and runs inference on embedded weights in milliseconds without needing to be copied or installed. What makes that possible is mmap(). Both the llama.cpp executable and the weights are concatenated onto the shell script. A tiny loader program is then extracted by the shell script, which maps the executable into memory. The llama.cpp executable then opens the shell script again as a file, and calls mmap() again to pull the weights into memory and make them directly accessible to both the CPU and GPU.

The trick to embedding weights inside llama.cpp executables is to ensure the local file is aligned on a page size boundary. That way, assuming the zip file is uncompressed, once it's mmap()'d into memory we can pass pointers directly to GPUs like Apple Metal, which require that data be page size aligned. Since no existing ZIP archiving tool has an alignment flag, we had to write about 400 lines of code to insert the ZIP files ourselves. However, once there, every existing ZIP program should be able to read them, provided they support ZIP64. This makes the weights much more easily accessible than they otherwise would have been, had we invented our own file format for concatenated files.

On Intel and AMD microprocessors, llama.cpp spends most of its time in the matmul quants, which are usually written thrice for SSSE3, AVX, and AVX2. llamafile pulls each of these functions out into a separate file that can be #includeed multiple times, with varying __attribute__((__target__("arch"))) function attributes. Then, a wrapper function is added which uses Cosmopolitan's X86_HAVE(FOO) feature to runtime dispatch to the appropriate implementation.

llamafile solves architecture portability by building llama.cpp twice: once for AMD64 and again for ARM64. It then wraps them with a shell script which has an MZ prefix. On Windows, it'll run as a native binary. On Linux, it'll extract a small 8kb executable called APE Loader to ${TMPDIR:-${HOME:-.}}/.ape that'll map the binary portions of the shell script into memory. It's possible to avoid this process by running the assimilate program that comes included with the cosmocc compiler. What the assimilate program does is turn the shell script executable into the host platform's native executable format. This guarantees a fallback path exists for traditonal release processes when it's needed.

Cosmopolitan Libc uses static linking, since that's the only way to get the same executable to run on six OSes. This presents a challenge for llama.cpp, because it's not possible to statically link GPU support. The way we solve that is by checking if a compiler is installed on the host system. For Apple, that would be Xcode, and for other platforms, that would be nvcc. llama.cpp has a single file implementation of each GPU module, named ggml-metal.m (Objective C) and ggml-cuda.cu (Nvidia C). llamafile embeds those source files within the zip archive and asks the platform compiler to build them at runtime, targeting the native GPU microarchitecture. If it works, then it's linked with platform C library dlopen() implementation. See llamafile/cuda.c and llamafile/metal.c.

In order to use the platform-specific dlopen() function, we need to ask the platform-specific compiler to build a small executable that exposes these interfaces. On ELF platforms, Cosmopolitan Libc maps this helper executable into memory along with the platform's ELF interpreter. The platform C library then takes care of linking all the GPU libraries, and then runs the helper program which longjmp()'s back into Cosmopolitan. The executable program is now in a weird hybrid state where two separate C libraries exist which have different ABIs. For example, thread local storage works differently on each operating system, and programs will crash if the TLS register doesn't point to the appropriate memory. The way Cosmopolitan Libc solves that is by JITing a trampoline around each dlsym() import, which blocks signals using sigprocmask() and changes the TLS register using arch_prctl(). Under normal circumstances, aspecting each function call with four additional system calls would be prohibitively expensive, but for llama.cpp that cost is infinitesimal compared to the amount of compute used for LLM inference. Our technique has no noticeable slowdown. The major tradeoff is that, right now, you can't pass callback pointers to the dlopen()'d module. Only one such function needed to be removed from the llama.cpp codebase, which was an API intended for customizing logging. In the future, Cosmoplitan will just trampoline signal handlers and code morph the TLS instructions to avoid these tradeoffs entirely. See cosmopolitan/dlopen.c for further details.

While the llamafile project is Apache 2.0-licensed, our changes to llama.cpp are licensed under MIT (just like the llama.cpp project itself) so as to remain compatible and upstreamable in the future, should that be desired.

The llamafile logo on this page was generated with the assistance of DALL·E 3.

• The 64-bit version of Windows has a 2GB file size limit. While llamafile will work fine on 64-bit Windows with the weights as a separate file, you'll get an error if you load them into the executable itself and try to run it.