Aira is a WebGPU-based artificial intelligence library built on a GPT-style architecture, written completely from scratch.
Aira.js is public now.Repo:Aira-prj/Aira.js
By running on the GPU, it delivers performance optimizations for tensor operations and training loops.
- Intel® HD Graphics 630:
- 221 lines dataset, 128 hidden, 64 embedSize, 3 epochs, learning rate 0.003
- Nvidia GeForce GTX 1050 Mobile:
- 221 lines dataset, 128 hidden, 64 embedSize, 3 epochs, learning rate 0.003
MLP Layer: Multilayer perceptron with forward and backward propagation
MHA (Multi-Head Attention): Transformer-based attention mechanism
LayerNorm: Normalization for training stability
Tensor Operations: GPU-accelerated matrix and vector multiplications
NLP Modules: BPE tokenization with encode/decode support
Training Loop: Training an MLP model on data with loss tracking
Sampling: Top-K and Top-P sampling, repeat penalty, and temperature control
Minimum(for testing and experimentation):
- CPU: Modern x64 CPU
- GPU: Intel HD Graphics 630 or equivalent
- RAM: 8 GB
- VRAM: ~1.5 GB (shared or dedicated)
Recommended:
- GPU: NVIDIA GTX 1650 / RTX 2080 or better
- Ram: 16-32 gb
- VRAM: 8–16 GB
Contact me:aira.project@proton.me
<script type="module" src="src/Aira.js">
(async () => {
const device = await ensureGPU();
const embedSize = 128
const hiddenSize = 512
const trainedModel = await StartModel("http://127.0.0.1:5500/datasets/en_Aira_big_chat_ds.txt", 500, embedSize, hiddenSize, 5, 0.003);
const usrprompt = "Hello!";
const mha = new MHA(device,embedSize,3)
const layernorm = new LayerNorm(embedSize)
const response = await generateReply(
mha,
trainedModel.mlp,
layernorm,
trainedModel.nlp,
trainedModel.embed,
usrprompt
);
console.log("Aira:", response);
})();
</script>