anudit/gpt-s2-5m-wgsl

GPT-S2-5M — WebGPU / WGSL assets

Runtime assets for running AxiomicLabs/GPT-S2-5M (a 9-layer custom GPTS3 language model) entirely in the browser with hand-written WebGPU (WGSL) compute shaders — no onnxruntime, no WASM inference.

The browser app fetches these files directly from this repo and caches them locally (Cache API), so the ~25 MB download happens only once.

Files

file	used by	notes
weights.bin	WGSL engine	all tensors as flat little-endian f32
manifest.json	WGSL engine	{name: {offset, shape}} index into weights.bin + model config
tokenizer.json, tokenizer_config.json, special_tokens_map.json	tokenizer	ByteLevel BPE, vocab 4096
generation_config.json	reference	bos/eos/pad ids
model.onnx, model.onnx.data	reference / parity	cache-free full-sequence ONNX export (not used by the WGSL engine)

weights.bin + manifest.json are produced from the source safetensors by scripts/export_weights.py; the ONNX pair by scripts/export_onnx.py.

Use

const BASE = "https://huggingface.co/anudit/gpt-s2-5m-wgsl/resolve/main";
const manifest = await (await fetch(`${BASE}/manifest.json`)).json();
const weights  = await (await fetch(`${BASE}/weights.bin`)).arrayBuffer();
// slice tensors out of `weights` by manifest offsets, upload to WebGPU, run WGSL.

See the source project for the full engine (src/gpts3/*).