warp-lang

Source code: https://github.com/NVIDIA/warp

Select version:

pip install warp-lang --index-url https://pypi.riseproject.dev/simple/

License: Apache-2.0

Download files: warp-lang-v1.17.0-20260929005900

Sources of bundled GPL libraries (gcc)

Patch applied for this version: https://github.com/riseproject-dev/python-wheels/tree/main/patches/warp-lang/1.17.0

pip install warp-lang==1.16.0 --index-url https://pypi.riseproject.dev/simple/

License: Apache-2.0

Download files: warp-lang-v1.16.0-20260925180804

Sources of bundled GPL libraries (gcc)

Patch applied for this version: https://github.com/riseproject-dev/python-wheels/tree/main/patches/warp-lang/1.16.0

Built in upstream’s own CPU-only configuration (build_lib.py --no-cuda), the one it uses for macOS: kernels are JIT-compiled to native riscv64 code by the bundled Clang/LLVM, and wp.get_devices() reports the CPU device only. The patches add riscv64 to the build and packaging platform tables, name the rv64gc/lp64d ABI for the JIT (which the Clang frontend does not default to), convert fp16 in software because rv64gc has no half-precision hardware and the JIT resolves no compiler-rt builtins, and let the environment raise both of the test runner’s hard-coded wall-clock limits (the parallel run’s overall cap and the tile-OOB tests’ per-subprocess cap).