W2GPU: Toward WebAssembly-to-WebGPU Program Translation via Small Language Models
summary
W2GPU fine-tunes small language models to translate WebAssembly Text to WGSL and produces a parallel dataset of 14,547 (WAT, WGSL) pairs derived from GLSL. Experiments use sub-2B models with 32K context (notably Qwen3 variants), show training loss convergence, end-to-end compilation success and computation result parity on unseen short kernels, but models fail to generalize to medium/long WAT sequences; the prototype is open source and works for small kernels.
AI-written summary · about
chapters (9)
- 0:00W2GPU project intro
- 2:30Abstraction paradox in programming
- 5:17W2GPU approach and training
- 7:47Challenges with long sequences
- 9:59Qualitative results and failures
- 10:40Traditional to ML-based compilation
- 11:42Empirical programming languages and LMPL
- 12:45Practical impact and open-source plans
- 14:09Q&A
related talks
30mWasi-Webgpu: Build Beautiful Graphics and Safely Run AI with Any GPU
33mTranslate tech talks using open-source models and cloud-native Wasm
32mEmpower WebAssembly ML Over Diverse AI Hardware Accelerators Outside the Browser
37mGetting Started with AI and WebAssembly
14mWasm, WebGPU, & WebNN: How compute abstraction are enabling client side AI
27mWasm Is Becoming the Runtime for LLMs
01h 42mWebAssembly: Writing Once and Running Everywhere NG
08mLightning Talk: Accelerate WebAssembly AI with Backend Graph Compliers on a Wide Variety of AI HW Accelerators