mrfakename/ncm-webgpu-weights

NCM small models, ONNX for the browser

ONNX exports of a few of the small Next-Concept Modeling checkpoints by Hazy Research (hazyresearch/ncm-checkpoints, code: HazyResearch/ncm, MIT). All credit for the models goes to the authors. They are used by the in-browser demo at mrfakename/ncm-webgpu.

Folder	Source checkpoint
tinystories/	tinystories
lm1b-small/	lm1b/small
sudoku-easy/, sudoku-medium/, sudoku-hard/	sudoku/{easy,medium,hard}

Each model.onnx (+ model.onnx.data, fp32) is one graph that runs the full coarse-to-fine generation, every level unrolled with shared weights, followed by the VQ-VAE decoder read-out of each level.

• Inputs: prefix (int64, 1 x prefix_len). The sampled models also take l0 (level-0 codes, drawn from level0_unigram.json), u (one uniform in [0, 1) per generated code) and temperature. Codes are drawn by inverse-CDF on softmax(logits / temperature), which matches torch.multinomial with the released top_k=0, top_p=1 settings.
• Outputs: codes, all levels concatenated, plus the decoded tokens of every level. For lm1b-small this is the top-128 logits per position, so the browser can run the released frequency and adjacency penalties (token_unigram.json, meta.json).
• meta.json has the level sizes and special tokens. fixtures.json has reference inputs and outputs from PyTorch. verification.json compares ONNX Runtime (CPU) with PyTorch.

Exported with torch.onnx.export (opset 17). Linear and area interpolation are rewritten as constant matrix products, which is the same linear map.