Host-only LLVM
LLVM and MLIR archives are consumed while linking the side module and are not part of the download. AMD bitcode helpers such asocml.bc,ockl.bc, andasanrtl.bcship as compiler inputs inside the package.
Triton · Pyodide · Emscripten
Build the compiler.
A compilation-only WebAssembly port of Triton for NVIDIA PTX, AMDGCN, and Gluon gfx950. The browser loads Python plus one side module; it never receives LLVM archives or a GPU runtime.
O4 + full LTO shipped
Current shipped profile
Full LTO · Release O3 · wasm-opt O4The deployed side module is 96,531,399 bytes after Full LTO, section garbage collection, andwasm-opt -O4. It passes the NVIDIA PTX, AMD gfx942, Gluon GFX950, and supplied v9 compilation-only checks. The older 92,972,220-byte O3 module is kept as the benchmark baseline; this profile favors compiler speed and backend coverage.
compile-only
What the build proves
Textual GPU output only; no driver or device is needed.
| Path | Target | Output | Boundary |
|---|---|---|---|
| Triton / NVIDIA | cuda:80:32 | PTX text | No cubin or ptxas |
| Triton / AMD | hip:gfx942:64 | AMDGCN text | No HSACO or ROCm |
| Gluon / AMD | hip:gfx950:64 | AMDGCN text | No GPU launch |
| Supplied v9 kernel | hip:gfx950:64 | LLIR + AMDGPU asm | Compile-only |
Stage 01
Pin compatible sources
Triton and LLVM are built from revisions known to agree. Only MLIR plus the WebAssembly, NVPTX, and AMDGPU targets are retained for the browser package.
Stage 02
Compile LLVM as a host library
LLVM and MLIR are linked once while producing the side module. Their static archives stay on the build machine; the browser never downloads them.
Stage 03
Package for Pyodide
The runtime payload is Python modules plus one Emscripten side module. CUDA, ROCm, ptxas, HSACO, and a GPU are intentionally absent.
Stage 04
Optimize and verify
The side module is linked with Full LTO, optimized with Binaryen at O4, then checked against every compilation-only target before it ships.
Compiler boundaries
These are the details that make the browser package different from native Triton.
GFX950 contract
The supplied Gluon kernel useship:gfx950:64, four warps,amdgpu-agpr-alloc=256, 16-byte pointer alignment, and K-contiguous B strides:stride_bk=1,stride_bn=K.
Inspect the output
Compiles emit TTGIR, LLIR, and AMDGPU assembly. The final assembly is textual AMDGCN; it is not HSACO and is not launchable by this package.