WebGPU backend (M10, phases 0-4) - #1

Merged

yhirose merged 12 commits into

worktree-webgpu-spike

Jul 20, 2026

Commits on Jul 20, 2026

ci: write the workflow comments in English
yhirose
committed
4c2e219
docs: stop tracking the internal design notes
yhirose
committed
35f076a
spike: WebGPU backend probe — sync flush() holds, WGSL sgemm ~75x CPU
yhirose
committed
96f8b0e
refactor: move the tl::gpu facade out of cuda.h into gpu.h
yhirose
committed
6598d6b
webgpu: host layer + WGSL sgemm, validated in the browser (M10 phases 1-2)
yhirose
committed
83776a6
webgpu: elementwise, broadcast and row reductions (M10 phase 3)
yhirose
committed
f027a2c
webgpu: calibrate the auto-mode thresholds (M10 phase 4)
yhirose
committed
2ce5b7a
webgpu: test the wasm backend on the same terms as CUDA
yhirose
committed
4ad404a
webgpu: run the wasm suite in CI, via Deno instead of a browser
yhirose
committed
d31a025
webgpu: fix a silent stale-readback, and batch the whole flush in one pass
yhirose
committed
cf1e906
docs: README had no idea the WebGPU backend exists
yhirose
committed
962fcbd
docs: add an API reference to the README
yhirose
committed
dc36af1