juntaoyuan/llawa · Hugging Face

Welcome to llawa

a.k.a Llama2 + Wasm QA

The models in this repo are Llama2 7b chat models further fine-tuned with Wasm-related Q&As. Instead of struggling with Python and PyTorch, the simplest way to run them on your own laptops, servers, or edge devices is to use the WasmEdge Runtime. Learn more about this fast, lightweight, portable, and ZERO Python dependency approach for running AI applications!

Install WasmEdge

curl -sSf https://raw.githubusercontent.com/WasmEdge/WasmEdge/master/utils/install.sh | bash -s -- --plugins wasi_nn-ggml

Download the chat app. It is a portal Wasm bytecode app. It runs across CPUs, GPUs, and OSes.

curl -LO https://github.com/second-state/llama-utils/raw/main/chat/llama-chat.wasm

Chat with the model on CLI

wasmedge --dir .:. \
  --nn-preload default:GGML:AUTO:llama-2-7b-chat-wasm-q5_k_m.gguf \
  llama-chat.wasm --reverse-prompt "</s>"

Below is an example chat session:

[USER]:
What is Wasm?
[ASSISTANT]:
Wasm is a binary instruction format for a stack-based virtual machine. It's a portable target for the WebAssembly runtime.
[USER]:
What is WASI-NN?
[ASSISTANT]:
WASI-NN is a Wasm system interface for neural networks. It allows Wasm apps to use and integrate with different neural network frameworks.
[USER]:
Which runtimes support it?
[ASSISTANT]:
Wasm is supported by a number of runtimes including WasmEdge, Wasmtime, and Wasmer.
[USER]:
How do I run AI workloads in serverless functions?
[ASSISTANT]:
You can use Wasm to run AI workloads in serverless functions. WasmEdge supports running AI workloads using the WASI-NN interface.

The llama-2-7b-chat-wasm-overfit-q5_k_m.gguf file is the fine-tuned model at epoch 25. It has a training loss of 0.03, and is probably over-fitted. You can try the above questions and see it give poor answers. We believe that training loss at 0.05 to 0.1 is optimal for this model.