suggest add numpy to dependance of codon-jit so that it can be installed automatically #808

Activity

1. BI71317 commented on May 15, 2026

   

   BI71317

   on May 15, 2026

   Contributor

   More actions

   MRE

   import codon
   from time import time
   
   
   @codon.jit(debug=True)
   def codon_numpy_sum(n: int) -> float:
       import numpy as np
   
       arr = np.arange(n, dtype=np.float64)
       return ((arr * 2.0) + 1.0).sum()
   
   
   def main() -> None:
       n = 1_000_000
   
       t0 = time()
       ans = codon_numpy_sum(n)
       t1 = time()
       print(f"[codon numpy only] result={ans:.1f} time={t1 - t0:.6f}s")
   
   
   if __name__ == "__main__":
       main()
   

   Result

   After uninstalling CPython NumPy:

   $ python3 -m pip uninstall numpy
   Found existing installation: numpy 2.4.4
   Uninstalling numpy-2.4.4:
     Would remove:
       /home/swchoi/.venvs/codon311/bin/f2py
       /home/swchoi/.venvs/codon311/bin/numpy-config
       /home/swchoi/.venvs/codon311/lib/python3.11/site-packages/numpy-2.4.4.dist-info/*
       /home/swchoi/.venvs/codon311/lib/python3.11/site-packages/numpy.libs/libgfortran-040039e1-0352e75f.so.5.0.0
       /home/swchoi/.venvs/codon311/lib/python3.11/site-packages/numpy.libs/libquadmath-96973f99-934c22de.so.0.0.0
       /home/swchoi/.venvs/codon311/lib/python3.11/site-packages/numpy.libs/libscipy_openblas64_-32a4b2a6.so
       /home/swchoi/.venvs/codon311/lib/python3.11/site-packages/numpy/*
   Proceed (Y/n)? Y
     Successfully uninstalled numpy-2.4.4
   

   Running the MRE fails before the Codon-compiled function body is actually used:

   $ python3 codon_numpy_only_demo.py 
   Traceback (most recent call last):
     File "/home/swchoi/src/test_code/python_codon_dsl/codon_numpy_only_demo.py", line 7, in <module>
       import codon
     File "/home/swchoi/src/codon/install/python/codon/__init__.py", line 7, in <module>
       from .decorator import jit, convert, execute, JITError, JITWrapper, _jit_register_fn, _jit_callback_fn, _jit
     File "/home/swchoi/src/codon/install/python/codon/decorator.py", line 13, in <module>
       import numpy as np
   ModuleNotFoundError: No module named 'numpy'
   (codon311) swchoi@swc-WIN:~/src/test_code/python_codon_dsl$ ```
   Running the MRE fails before the Codon-compiled function body is actually used:
   

   Root Cause

   This failure happens even when the user code only intends to use Codon's native NumPy support inside the JIT-compiled function, not CPython's NumPy module directly.

   Looking at codon/install/python/codon/decorator.py, the decorator currently refers to CPython NumPy scalar types:

   pod_conversions = {
       type(None): "pyobj",
       int: "int",
       float: "float",
       bool: "bool",
       str: "str",
       complex: "complex",
       slice: "slice",
       np.bool_: "bool",
       np.int8: "i8",
       np.uint8: "u8",
       np.int16: "i16",
       np.uint16: "u16",
       np.int32: "i32",
       np.uint32: "u32",
       np.int64: "int",
       np.uint64: "u64",
       np.float16: "float16",
       np.float32: "float32",
       np.float64: "float",
       np.complex64: "complex64",
       np.complex128: "complex",
   }
   

   Because of this, importing codon through the Python JIT path implicitly depends on the CPython NumPy package being installed,

   even if Codon Native NumPy is only imported inside the Codon-compiled function body.

   My understanding is that these entries are used to map Python / NumPy scalar types to Codon's statically typed representation.

   Since Python's builtin numeric types do not always encode the exact fixed-width type information needed here, NumPy scalar types are being used as convenient concrete type markers.

   A possible direction may be to avoid requiring CPython NumPy at module import time by using another internal/builtin representation for fixed-width scalar type markers.

   Alternatively, if this dependency is intentional, it may be worth documenting that codon-jit currently requires CPython NumPy for NumPy scalar type conversion support, or just add module dependency,

   even when the compiled function body uses Codon's native NumPy implementation.