saksornr/khammoon-webgpu-q4

Khammoon WebGPU Q4

Independent browser conversion of typhoon-ai/khammoon, pinned at revision 52e6550d3cd41d3a22d0649cb06bcdb67b3bef46.

Used by the Typhoon WebGPU Playground. Model files are downloaded on demand, and neural inference runs on the visitor’s WebGPU hardware. Input text, images, and audio stay in the browser.

Model and conversion

Thai morpheme segmentation using the exact Khammoon character-CNN and two-layer bidirectional LSTM. It splits text into meaningful Thai units, not LLM vocabulary tokens. The browser implementation ports the upstream character classes, Thai Character Cluster boundary rules, newline context, and overlapping 256-character windows.

All embedding, convolution, recurrent, and classifier matrices are packed as symmetric block-32 INT4. Activations, biases, and quantization scales remain FP32. Neural operations run in native WebGPU kernels; character encoding, window assembly, and boundary decoding run in JavaScript. Original characters and whitespace are preserved.

The package contains weights.bin, scales.bin, biases.bin, vocab.json, and manifest.json. Use the application’s Khammoon runtime. Hardware-GPU smoke tests matched the upstream sample tokenizations, and numerical comparisons against dequantized-Q4 NumPy weights stayed below 5e-7 maximum probability error. These small checks are not a corpus accuracy benchmark.

Browser runtime

Custom native WebGPU WGSL kernels; this package is not an AutoTokenizer or generic ONNX model.

Use a current desktop Chrome or Edge with hardware acceleration enabled. Tokenization/preprocessing/decoding can run in JavaScript; neural inference uses WebGPU. The application selects no CPU model or full-precision model fallback. Four-bit refers to the quantized weights; the floating-point exceptions are described above.

All package files are at this repository’s root, with ONNX graphs under onnx/ where applicable. manifest.json records source provenance, quantization, file sizes, and SHA-256 checksums. Exporters and runtime integration are available in the playground’s source repository. This is an independent conversion, not an official Typhoon release, and attribution does not imply endorsement.

License and attribution

Weights use CC BY-NC-SA 4.0, including noncommercial and share-alike terms. See NOTICE for attribution and conversion details. The upstream Python/TCC code is Apache-2.0; that does not change the weights’ noncommercial license.