1 unstable release

0.1.0 Mar 2, 2024

#630 in Machine learning

24 downloads per month
Used in 2 crates

Apache-2.0

170KB
4K SLoC

crabml

crabml is a llama.cpp compatible (and equally fast!) AI inference engine written in 🦀 Rust, which runs everywhere with the help of 🎮 WebGPU.

Project Goals

crabml is designed with the following objectives in mind:

  • 🤖 Focus solely on inference.
  • 🎮 Runs on browsers, desktops, and servers everywhere with the help of WebGPU.
  • SIMD-accelerated inference on inexpensive hardware.
  • 💼 mmap() from day one, minimized memory requirement with various quantization support.
  • 👾 Hackable & embeddable.

Supported Models

crabml supports the following models in GGUF format:

  • 🦙 Llama
  • 🦙 CodeLlama
  • 🦙 Gemma
  • 〽️ Mistral
  • 🚄 On the way: Mistral MoE, Phi, QWen, StarCoder, Llava, and more!

For more information, you can visit How to Get GGUF Models to learn how to download the GGUF files you need.

Supported Quantization Methods

crabml supports the following quantization methods on CPUs with SIMD acceleration for ARM (including Apple Silicon) and x86 architectures:

Bits Native CPU NEON AVX2 RISC-V SIMD WebGPU
Q8_0 8 bits WIP WIP
Q6_K 6 bits WIP WIP WIP WIP
Q5_0 5 bits WIP WIP WIP WIP
Q5_1 5 bits WIP WIP WIP WIP
Q5_K 5 bits WIP WIP WIP WIP
Q4_0 4 bits WIP WIP WIP
Q4_1 4 bits WIP WIP
Q4_K 4 bits WIP WIP WIP WIP
Q3_K 3 bits WIP WIP WIP WIP
Q2_K 2 bits WIP WIP WIP WIP

As the table above suggests, WebGPU-accelerated quantizations are still under busy development, and Q8_0Q4_0Q4_1 are currently the most recommended quantization methods on CPUs!

Usage

Building the Project

To build crabml, set the RUSTFLAGS environment variable to enable specific target features. For example, to enable NEON on ARM architectures, use RUSTFLAGS="-C target-feature=+neon". Then build the project with the following command:

cargo build --release

This command compiles the project in release mode, which optimizes the binary for performance.

Running an Example

After building the project, you can run an example inference by executing the crabml-cli binary with appropriate arguments. For instance, to use the tinyllamas-stories-15m-f32.gguf model to generate text based on the prompt "captain america", execute the command below:

./target/release/crabml-cli \
  -m ./testdata/tinyllamas-stories-15m-f32.gguf \
  "captain america" --steps 100 \
  -t 0.8 -p 1.0

In this command:

  • -m specifies the checkpoint file.
  • --steps defines the number of tokens to generate.
  • -t sets the temperature, which controls the randomness of the output.
  • -p sets the probability of sampling from the top-p.

License

This contribution is licensed under Apache License, Version 2.0, (LICENSE or http://www.apache.org/licenses/LICENSE-2.0)

Dependencies

~9–45MB
~676K SLoC