| license: apache-2.0 | |
| base_model: convaiinnovations/laya | |
| library_name: onnx | |
| tags: [onnx, onnxruntime-web, webgpu, laya] | |
| # Laya for WebGPU in the browser (int8 ONNX, fp32 embeddings) | |
| Unofficial copy of [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) (Apache-2.0, | |
| Copyright ConvAI Innovations) for onnxruntime-web. Used by https://archevel.github.io/laya/. | |
| Made from [harshpreet931/cut-laya-onnx](https://huggingface.co/harshpreet931/cut-laya-onnx) | |
| (revision `192138dfb32a0b2f73425ba70345946d415bfde1`: 8-bit weight-only MatMulNBits, split encoder/head for | |
| [laya-ts](https://github.com/NandhaKishorM/laya/tree/main/laya-ts)) with one change: the token-embedding table in | |
| `encoder.onnx` is widened from fp16 to fp32 (lossless). With the fp16 table, onnxruntime-web's WebGPU Gather | |
| requires the `shader-f16` feature, which Chrome does not expose on e.g. Linux with NVIDIA GPUs. | |
| Script: `tools/embeddings_to_fp32.py` in https://github.com/archevel/laya. | |
| Outputs differ slightly from the PyTorch checkpoint because of the 8-bit weights. | |
| Licensed under the Apache License, Version 2.0, like the original. | |