tokenspeed-deepjit

Python packaging for the header-only deepseek-ai/DeepJIT C++20 library. The tokenspeed-deepjit distribution vendors the headers from upstream commit 8b3ef868705a3792cc1a14ab570b539f72fd3d94 and provides Python and CMake helpers for locating them.

Installation

pip install tokenspeed-deepjit

The wheel is platform-independent because DeepJIT is header-only. Applications still need the device toolchain and libraries required by the selected backend. CUDA consumers need CUDA headers 12.4 or newer, NVCC 12.9 or newer, and PyTorch with CUDA support. Ascend consumers need the CANN compiler and headers, ACL, and torch_npu. Pybind11 is also required when exposing DeepJIT through a Python extension.

Python helpers

import tokenspeed_deepjit

print(tokenspeed_deepjit.include_dir())
print(tokenspeed_deepjit.cmake_prefix_path())

The distribution does not install a deep_jit Python import package. The upstream import package is empty; omitting it avoids conflicts with extensions that embed DeepJIT.

CMake usage

find_package(Python COMPONENTS Interpreter REQUIRED)
execute_process(
  COMMAND "${Python_EXECUTABLE}" -c
          "import tokenspeed_deepjit; print(tokenspeed_deepjit.cmake_prefix_path())"
  OUTPUT_VARIABLE tokenspeed_deepjit_ROOT
  OUTPUT_STRIP_TRAILING_WHITESPACE
  COMMAND_ERROR_IS_FATAL ANY
)

list(PREPEND CMAKE_PREFIX_PATH "${tokenspeed_deepjit_ROOT}")
find_package(deep_jit CONFIG REQUIRED)
target_link_libraries(my_target PRIVATE deep_jit::deep_jit)

The imported target requests C++20 for C++ sources. CUDA projects must also set CUDA_STANDARD 20 or pass the equivalent standard flag to NVCC.

Local build

Populate the upstream headers before building:

git clone https://github.com/deepseek-ai/DeepJIT.git /tmp/DeepJIT
git -C /tmp/DeepJIT checkout 8b3ef868705a3792cc1a14ab570b539f72fd3d94
./prepare_headers.sh /tmp/DeepJIT
python -m build

See the upstream documentation for the runtime API and backend integration details.