cjpais/whisperfile

Notice

This code has been merged into llamafile. All future work will be done in that repo!

The original README for the project is below.

whisperfile

This is a fork of llamafile which builds llamafiles for whisper.cpp. A huge credit and thanks to the original authors of these wonderful projects.

Much of this README will be a copy from the original project, with some modifications to reflect the changes made in this fork.

NOTE: Currently this project only supports the server binary of whisper.cpp. In the future there may be support for the other binaries as well, or a wrapper binary similar to the original llamafile project to support both cli and server implementations.

Quickstart

You can find prebuilt whisperfiles here: https://huggingface.co/cjpais/whisperfile

To run the whisperfile, download it and run it like so:

chmod +x whisper.base.llamafile
./whisper.base.llamafile

This will start a whisper.cpp server at port 51524. You can access the server by visiting http://localhost:51524 in your browser. It is set up to automatically convert audio files into the correct format with the --convert flag. This requires ffmpeg to be installed on the host system. The API endpoint is the same as the original whisper.cpp server. You can find the documentation for it at here

How llamafile works

A llamafile is an executable LLM that you can run on your own computer. It contains the weights for a given open LLM, as well as everything needed to actually run that model on your computer. There's nothing to install or configure (with a few caveats, discussed in subsequent sections of this document).

This is all accomplished by combining llama.cpp with Cosmopolitan Libc, which provides some useful capabilities:

1. 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.

2. 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.

3. llamafiles can run on six OSes (macOS, Windows, Linux, FreeBSD, OpenBSD, and NetBSD). If you make your own llama files, 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.

4. The weights for an LLM can be embedded within the 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.

5. Finally, with the tools included in this project you can create your own llamafiles, using any compatible model weights you want. You can then distribute these llamafiles to other people, who can easily make use of them regardless of what kind of computer they have.

Gotchas

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

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 some Linux systems, you might get errors relating to run-detectors or WINE. This is due to binfmt_misc registrations. You can fix that by adding an additional registration for the APE file format llamafile uses:

sudo wget -O /usr/bin/ape https://cosmo.zip/pub/cosmos/bin/ape-$(uname -m).elf
sudo chmod +x /usr/bin/ape
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"

As mentioned above, on Windows you may need to rename your llamafile by adding .exe to the filename.

Also as mentioned above, Windows also has a maximum file size limit of 4GB for executables. The LLaVA server executable above is just 30MB shy of that limit, so it'll work on Windows, but with larger models like WizardCoder 13B, you need to store the weights in a separate file. An example is provided above; see "Using llamafile with external weights."

On WSL, it's recommended that the WIN32 interop feature be disabled:

sudo sh -c "echo -1 > /proc/sys/fs/binfmt_misc/WSLInterop"

In the instance of getting a Permission Denied on disabling interop through CLI, it can be permanently disabled by adding the following in /etc/wsl.conf

[interop]
enabled=false

On Raspberry Pi, if you get "mmap error 12" then it means your kernel is configured with fewer than 48 bits of address space. You need to upgrade to RPI 5. You can still use RPI 4 if you either (1) rebuild your kernel, or (2) get your SDcard OS image directly from Ubuntu (don't use RPI OS).

On any platform, if your llamafile process is immediately killed, check if you have CrowdStrike and then ask to be whitelisted.

Supported OSes

llamafile supports the following operating systems, which require a minimum stock install:

• Linux 2.6.18+ (i.e. every distro since RHEL5 c. 2007)
• Darwin (macOS) 23.1.0+ [1] (GPU is only supported on ARM64)
• Windows 8+ (AMD64 only)
• FreeBSD 13+
• NetBSD 9.2+ (AMD64 only)
• OpenBSD 7+ (AMD64 only)

On Windows, llamafile runs as a native portable executable. On UNIX systems, llamafile extracts a small loader program named ape to $TMPDIR/.llamafile or ~/.ape-1.9 which is used to map your model into memory.

[1] Darwin kernel versions 15.6+ should be supported, but we currently have no way of testing that.

Supported CPUs

llamafile supports the following CPUs:

• AMD64 microprocessors must have AVX. Otherwise llamafile will print an error and refuse to run. This means that if you have an Intel CPU, it needs to be Intel Sandybridge or newer (circa 2011+), and if you have an AMD CPU, then it needs to be Bulldozer or newer (circa 2011+). Support for AVX512, AVX2, FMA, F16C, and VNNI are conditionally enabled at runtime if you have a newer CPU.

• ARM64 microprocessors must have ARMv8a+. This means everything from Apple Silicon to 64-bit Raspberry Pis will work, provided your weights fit into memory.

GPU support

llamafile supports the following kinds of GPUs:

• Apple Metal
• NVIDIA
• AMD

GPU on MacOS ARM64 is supported by compiling a small module using the Xcode Command Line Tools, which need to be installed. This is a one time cost that happens the first time you run your llamafile. The DSO built by llamafile is stored in $TMPDIR/.llamafile or $HOME/.llamafile. Offloading to GPU is enabled by default when a Metal GPU is present. This can be disabled by passing -ngl 0 or --gpu disable to force llamafile to perform CPU inference.

Owners of NVIDIA and AMD graphics cards need to pass the -ngl 999 flag to enable maximum offloading. If multiple GPUs are present then the work will be divided evenly among them by default, so you can load larger models. Multiple GPU support may be broken on AMD Radeon systems. If that happens to you, then use export HIP_VISIBLE_DEVICES=0 which forces llamafile to only use the first GPU.

Windows users are encouraged to use our release binaries, because they contain prebuilt DLLs for both NVIDIA and AMD graphics cards, which only depend on the graphics driver being installed. If llamafile detects that NVIDIA's CUDA SDK or AMD's ROCm HIP SDK are installed, then llamafile will try to build a faster DLL that uses cuBLAS or rocBLAS. In order for llamafile to successfully build a cuBLAS module, it needs to be run on the x64 MSVC command prompt. You can use CUDA via WSL by enabling Nvidia CUDA on WSL and running your llamafiles inside of WSL. Using WSL has the added benefit of letting you run llamafiles greater than 4GB on Windows.

On Linux, NVIDIA users will need to install the CUDA SDK (ideally using the shell script installer) and ROCm users need to install the HIP SDK. They're detected by looking to see if nvcc or hipcc are on the PATH.

If you have both an AMD GPU and an NVIDIA GPU in your machine, then you may need to qualify which one you want used, by passing either --gpu amd or --gpu nvidia.

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.

Source installation

Developing on llamafile requires a modern version of the GNU make command (called gmake on some systems), sha256sum (otherwise cc will be used to build it), wget (or curl), and unzip available at https://cosmo.zip/pub/cosmos/bin/. Windows users need cosmos bash shell too.

make -j8
sudo make install PREFIX=/usr/local

Creating whisperfiles

If you want to be able to just say:

./whisper-tiny.llamafile

...and have it run the web server without having to specify arguments, then you can embed both the weights and a special .args inside, which specifies the default arguments. First, let's create a file named .args which has this content:

-m
tiny.bin
--host
0.0.0.0
--port
51524
--convert
-pc
-pr
...

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 both the weights and the argument file to the executable:

cp /usr/local/bin/whisperfile whisper-tiny.llamafile

zipalign -j0 \
  whisper-tiny.llamafile \
  tiny.bin \
  .args

./llava.llamafile

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

Distribution

One good way to share a llamafile with your friends is by posting it on Hugging Face. If you do that, then it's recommended that you mention in your Hugging Face commit message what git revision or released version of llamafile you used when building your llamafile. That way everyone online will be able verify the provenance of its executable content. If you've made changes to the llama.cpp or cosmopolitan source code, then the Apache 2.0 license requires you to explain what changed. One way you can do that is by embedding a notice in your llamafile using zipalign that describes the changes, and mention it in your Hugging Face commit.

Documentation

There's a manual page for each of the llamafile programs installed when you run sudo make install. The command manuals are also typeset as PDF files that you can download from our GitHub releases page. Lastly, most commands will display that information when passing the --help flag.

Technical details

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.

ZIP weights embedding

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 500 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.

Microarchitectural portability

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.

Architecture portability

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 traditional release processes when it's needed.

GPU support

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 on AMD is by using SSE to recompile the executable at runtime to change %fs register accesses into %gs which takes a millisecond. On ARM, Cosmo uses the x28 register for TLS which can be made safe by passing the -ffixed-x28 flag when compiling GPU modules. Lastly, llamafile uses the __ms_abi__ attribute so that function pointers passed between the application and GPU modules conform to the Windows calling convention. Amazingly enough, every compiler we tested, including nvcc on Linux and even Objective-C on MacOS, all support compiling WIN32 style functions, thus ensuring your llamafile will be able to talk to Windows drivers, when it's run on Windows, without needing to be recompiled as a separate file for Windows. See cosmopolitan/dlopen.c for further details.

Security

TODO: Currently security with pledge is not implemented for whisperfile.

WARNING: Given this you may consider the security to be quite poor.

Building Release

I run the following commands:

./llamafile/cuda.sh
./llamafile/rocm.sh
zipalign -j0 o/whisper.cpp/server/server ~/ggml-cuda.so ~/ggml-rocm.so

Licensing

The whisperfile project is Apache 2.0-licensed. 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.