lazy import cuBLAS

cupy/_core/_routines_linalg.pyx

@@ -27,7 +27,6 @@ from cupy._core cimport _routines_manipulation as _manipulation
 from cupy._core cimport _routines_math as _math
 from cupy.cuda cimport device
 from cupy_backends.cuda.api cimport runtime
-from cupy_backends.cuda.libs cimport cublas
 
 
 cdef extern from '../../cupy_backends/cupy_complex.h':
@@ -543,6 +542,8 @@ cpdef _ndarray_base dot(
 cpdef _ndarray_base tensordot_core(
         _ndarray_base a, _ndarray_base b, _ndarray_base out, Py_ssize_t n,
         Py_ssize_t m, Py_ssize_t k, const shape_t& ret_shape):
+    from cupy_backends.cuda.libs import cublas
+
     # out, if specified, must be C-contiguous and have correct shape.
     cdef shape_t shape
     cdef Py_ssize_t transa, transb, lda, ldb
@@ -704,6 +705,8 @@ cpdef _ndarray_base tensordot_core_v11(
         Py_ssize_t transa, Py_ssize_t transb, Py_ssize_t m, Py_ssize_t n,
         Py_ssize_t k, _ndarray_base a, Py_ssize_t lda, _ndarray_base b,
         Py_ssize_t ldb, _ndarray_base c, Py_ssize_t ldc):
+    from cupy_backends.cuda.libs import cublas
+
     cdef float one_f, zero_f
     cdef double one_d, zero_d
     cdef cuComplex one_F, zero_F
@@ -833,6 +836,7 @@ cpdef _ndarray_base matmul(
     .. seealso:: :func:`numpy.matmul`
 
     """
+    from cupy_backends.cuda.libs import cublas
 
     cdef Py_ssize_t i, n, m, ka, kb, a_sh, b_sh, c_sh, ldc
     cdef Py_ssize_t batchCount, a_part_outshape, b_part_outshape

cupy/_core/core.pyx

@@ -48,7 +48,6 @@ from cupy.cuda cimport memory
 from cupy.cuda cimport stream as stream_module
 from cupy_backends.cuda cimport stream as _stream_module
 from cupy_backends.cuda.api cimport runtime
-from cupy_backends.cuda.libs cimport cublas
 
 
 # If rop of cupy.ndarray is called, cupy's op is the last chance.
@@ -2674,6 +2673,8 @@ cpdef _ndarray_base _internal_ascontiguousarray(_ndarray_base a):
 
 
 cpdef _ndarray_base _internal_asfortranarray(_ndarray_base a):
+    from cupy_backends.cuda.libs import cublas
+
     cdef _ndarray_base newarray
     cdef int m, n
     cdef intptr_t handle

cupy/cuda/__init__.py

@@ -17,7 +17,6 @@
 from cupy.cuda import texture  # NOQA
 from cupy_backends.cuda.api import driver  # NOQA
 from cupy_backends.cuda.api import runtime  # NOQA
-from cupy_backends.cuda.libs import cublas  # NOQA
 from cupy_backends.cuda.libs import nvrtc  # NOQA
 from cupy_backends.cuda.libs import profiler  # NOQA
 
@@ -63,6 +62,10 @@ def __getattr__(key):
         from cupy_backends.cuda.libs import curand
         _cupy.cuda.curand = curand
         return curand
+    elif key == 'cublas':
+        from cupy_backends.cuda.libs import cublas
+        _cupy.cuda.cublas = cublas
+        return cublas
 
     # `nvtx_enabled` flags are kept for backward compatibility with Chainer.
     # Note: module-level getattr only runs on Python 3.7+.

cupy/cuda/device.pyx

@@ -5,7 +5,6 @@ import threading
 from cupy._core import syncdetect
 from cupy_backends.cuda.api cimport runtime
 from cupy_backends.cuda.api import runtime as runtime_module
-from cupy_backends.cuda.libs import cublas
 from cupy import _util
 
 
@@ -251,6 +250,7 @@ cdef class Device:
         itself is different.
 
         """
+        from cupy_backends.cuda.libs import cublas
         return self._get_handle(
             'cublas_handles', cublas.create, cublas.destroy)
 

cupy/linalg/_decomposition.py

@@ -2,7 +2,6 @@
 
 import cupy
 from cupy_backends.cuda.api import runtime
-from cupy_backends.cuda.libs import cublas
 from cupy._core import internal
 from cupy.cuda import device
 from cupy.linalg import _util
@@ -33,6 +32,7 @@ def _lu_factor(a_t, dtype):
     .. seealso:: :func:`scipy.linalg.lu_factor`
 
     """
+    from cupy_backends.cuda.libs import cublas
     from cupy_backends.cuda.libs import cusolver
 
     orig_shape = a_t.shape
@@ -57,13 +57,13 @@ def _lu_factor(a_t, dtype):
         a_array = cupy.arange(start, stop, step, dtype=cupy.uintp)
 
         if dtype == numpy.float32:
-            getrfBatched = cupy.cuda.cublas.sgetrfBatched
+            getrfBatched = cublas.sgetrfBatched
         elif dtype == numpy.float64:
-            getrfBatched = cupy.cuda.cublas.dgetrfBatched
+            getrfBatched = cublas.dgetrfBatched
         elif dtype == numpy.complex64:
-            getrfBatched = cupy.cuda.cublas.cgetrfBatched
+            getrfBatched = cublas.cgetrfBatched
         elif dtype == numpy.complex128:
-            getrfBatched = cupy.cuda.cublas.zgetrfBatched
+            getrfBatched = cublas.zgetrfBatched
         else:
             assert False
 
@@ -117,8 +117,10 @@ def _potrf_batched(a):
     Returns:
         cupy.ndarray: The lower-triangular matrix.
     """
+    from cupy_backends.cuda.libs import cublas
     from cupy_backends.cuda.libs import cusolver
     from cupyx.cusolver import check_availability
+
     if not check_availability('potrfBatched'):
         raise RuntimeError('potrfBatched is not available')
 
@@ -175,6 +177,7 @@ def cholesky(a):
 
     .. seealso:: :func:`numpy.linalg.cholesky`
     """
+    from cupy_backends.cuda.libs import cublas
     from cupy_backends.cuda.libs import cusolver
 
     _util._assert_cupy_array(a)

cupy/linalg/_eigenvalue.py

@@ -1,7 +1,6 @@
 import numpy
 
 import cupy
-from cupy_backends.cuda.libs import cublas
 from cupy.cuda import device
 from cupy.cuda import runtime
 from cupy.linalg import _util
@@ -12,6 +11,7 @@
 
 
 def _syevd(a, UPLO, with_eigen_vector, overwrite_a=False):
+    from cupy_backends.cuda.libs import cublas
     from cupy_backends.cuda.libs import cusolver
 
     if UPLO not in ('L', 'U'):

cupy/linalg/_solve.py

@@ -8,7 +8,6 @@
 from cupy.cuda import device
 from cupy.linalg import _decomposition
 from cupy.linalg import _util
-from cupy.cublas import batched_gesv, get_batched_gesv_limit
 import cupyx
 
 
@@ -36,6 +35,8 @@ def solve(a, b):
 
     .. seealso:: :func:`numpy.linalg.solve`
     """
+    from cupy.cublas import batched_gesv, get_batched_gesv_limit
+
     if a.ndim > 2 and a.shape[-1] <= get_batched_gesv_limit():
         # Note: There is a low performance issue in batched_gesv when matrix is
         # large, so it is not used in such cases.

cupyx/__init__.py

@@ -9,7 +9,6 @@
 from cupyx import time  # NOQA
 from cupyx import scipy  # NOQA
 from cupyx import optimizing  # NOQA
-from cupyx import lapack  # NOQA
 
 from cupyx._ufunc_config import errstate  # NOQA
 from cupyx._ufunc_config import geterr  # NOQA

cupyx/linalg/_solve.py

@@ -1,5 +1,4 @@
 from cupy.linalg import _util
-from cupyx import lapack
 
 
 def invh(a):
@@ -15,6 +14,7 @@ def invh(a):
     Returns:
         cupy.ndarray: The inverse of matrix ``a``.
     """
+    from cupyx import lapack
 
     _util._assert_cupy_array(a)
     # TODO: Use `_assert_stacked_2d` instead, once cusolver supports nrhs > 1