| --- |
| license: mit |
| library_name: tinygrad |
| tags: |
| - gpt2 |
| - webgpu |
| - tinygrad |
| - logit-lens |
| --- |
| |
| # tinygpt-webgpu |
|
|
| GPT-2 124M compiled to WGSL compute kernels by [tinygrad]'s WebGPU exporter, for |
| running inference in the browser at <https://jik.wtf/tinygpt>. |
|
|
| This repository holds only the weight payloads. The compiled kernels and the |
| manifests that describe these blobs live in the site repository beside the |
| loader, because the offsets in a manifest are meaningful only against the exact |
| blob they were computed for. |
|
|
| | file | bytes | contents | |
| | --- | --- | --- | |
| | `weights-*.bin` | 495,399,936 | transformer weights, fp32, laid out for the exported graph | |
| | `lens-*.bin` | 30,670,848 | fitted J-lens Jacobians, 13 layers | |
|
|
| Weights are delivered as 16 MiB content-addressed chunks. Each file name carries |
| the SHA-256 of its own contents, so a chunk URL always refers to exactly one |
| sequence of bytes: safe to cache indefinitely, and impossible to pair with a |
| manifest describing a different packing. The manifests listing them live in the |
| site repository beside the loader. |
|
|
| The decode graph is compiled for a fixed 256-slot KV cache and emits both logits |
| and all 13 residual stream taps, so both lenses read from one forward pass. |
|
|
| Greedy decoding reproduces the native tinygrad oracle token for token: |
| `"The capital of France is"` continues `" the capital of the French"`. |
|
|
| [tinygrad]: https://github.com/tinygrad/tinygrad |
|
|
