C++ extension¶

torch_tvarant is a PyTorch C++ extension (PrivateUse1) built from csrc/.

Layout¶

Path	Role
`csrc/Module.cpp`	pybind entry: device helpers, force-link registrations
`csrc/aten/`	ATen `TORCH_LIBRARY_IMPL(aten, PrivateUse1, …)`
`csrc/runtime/`	CPU sim + OpenCL runtimes
`csrc/kernels/host/`	Host fp32 kernels
`csrc/kernels/opencl/`	OpenCL `.cl` sources
`csrc/jit/`	Pointwise / GEMM fusion helpers

Registering an ATen op¶

TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
  m.impl("relu", TORCH_FN(tvarant::ops::relu));
}

Implement the kernel (host + optional OpenCL), wire runtime().launch, then register the impl in Ops.cpp, LlmOps.cpp, or another *Ops.cpp TU that is force-linked from Module.cpp.

Qualify calls into tvarant::ops::… when names collide with ATen ADL (see existing silu / unary helpers).

Custom library¶

TORCH_LIBRARY(tvarant, m) in CustomOps.cpp defines torch.ops.tvarant.linear_act and torch.ops.tvarant.pointwise.

Build¶

pip install -e . --no-build-isolation
# OpenCL:
USE_OPENCL=1 pip install -e . --no-build-isolation

See Installation and Contributing.