Pass-trace compiler

Bring a kernel.

Trace any decorated Triton or Gluon kernel locally—no GPU launch. The AMD + NVIDIA Pyodide compiler runs in a dedicated browser worker.

Kernel configuration

Kernel preset

The GFX950 preset is the supplied CDNA4 kernel with compile-only specialization metadata.

Language

Entrypoint kernel name

Exact decorated function name; leading underscores are valid.

Compile signature

One item per parameter. Numeric values supplytl.constexprparameters.

Target

BACKEND:ARCH:WARP_SIZE · e.g. hip:gfx950:64

num_warps

num_stages

LLVM function attrs

The GFX950 preset usesamdgpu-agpr-alloc=256.

Pointer alignment

Addstt.divisibilityto pointer arguments.

Source

loading Python/Triton grammar + editor…

Paste any valid decorated Triton or Gluon kernel · 100 KB limit ·import torchis supported for compile-only source and annotations · fixed compiler invocation · isolated temporary directory