anudit/supra-1.5-50m-wgsl

Supra-1.5-50M-Instruct — WebGPU / WGSL (int8) assets

Runtime assets for running SupraLabs/Supra-1.5-50M-Instruct-exp (a 12-layer Llama, 50M params) entirely in the browser with hand-written WebGPU (WGSL) compute shaders — no onnxruntime, no WASM inference.

Weights are 8-bit (Q8): symmetric per-output-row int8 + one f32 scale per row, quantized from the original bf16 safetensors. Cosine similarity of the next-token logits vs full precision is 0.9997 (argmax unchanged). The browser fetches these files directly and caches them (Cache API), so the ~52 MB download happens once.

Files

file	notes
weights_q8.bin	packed int8 weights (row-major, 4 int8 per u32 for WGSL unpack4x8snorm) + per-row f32 scales; RMSNorm weights in f32
manifest.json	{name: {type, N, K, wOffset, sOffset}} index + model config
tokenizer.json, tokenizer_config.json	Llama tokenizer, vocab 32000
config.json, generation_config.json	architecture + bos/eos/pad ids

Produced by scripts/export_supra_q8.py. In-shader dequant: W ≈ scale127[n] * unpack4x8snorm(packed).

Architecture

Llama: hidden 512, 12 layers, 8 heads / 4 KV (GQA), head_dim 64, intermediate 1408 (SwiGLU), RoPE θ=10000 (rotate-half), RMSNorm, tied embeddings, vocab 32000.

Prompt format (Alpaca)

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response: