tvm_tirx.testing.env

Local runtime, device, and compiler capability probes.

Functions

has_cuda()	True if a CUDA device is present.
has_rocm()	True if a ROCm device is present.
has_vulkan()	True if a Vulkan device is present.
has_metal()	True if a Metal device is present.
has_opencl()	True if an OpenCL device is present.
has_nvptx()	True if NVPTX is usable: a (CUDA) device, plus the LLVM backend it needs.
has_llvm()	True if the LLVM backend was built in.
has_gpu()	True if any GPU backend (cuda/rocm/opencl/metal/vulkan) is present.
has_multi_gpu([count])	True if at least count devices of a single GPU backend exist.
cuda_arch([device_id])	Return a CUDA device's canonical runtime architecture, if available.
has_cuda_arch(*arches[, device_id])	True if a CUDA device exactly matches one of arches.
has_cuda_compute(major[, minor, exact])	True if the CUDA compute capability satisfies (major, minor).
has_nvcc_version(major[, minor, release])	True if a CUDA device is present and nvcc is at least (major, minor, release).
has_llvm_min_version(major)	True if LLVM is available and its major version is at least major.
has_matrixcore()	True if a ROCm device with Matrix Core support (compute >= 8) exists.
has_cudagraph()	True if a CUDA device is present and the toolkit supports CUDA Graphs.
has_cpu_feature(features)	True if the host CPU supports features (a name or list of names).
has_x86_vnni()	True if the host CPU supports x86 VNNI (AVX512-VNNI or AVX-VNNI).
has_x86_avx512()	True if the host CPU supports the x86 AVX512 extensions.