gpu4pyscf 1.7.2__tar.gz → 1.7.4__tar.gz

{gpu4pyscf-1.7.2 → gpu4pyscf-1.7.4}/CHANGELOG

@@ -1,3 +1,15 @@
+v1.7.4 (2026-06-23)
+-------------------
+* Improvements
+  - Check host memory limits before allocating large mapped memory
+
+
+v1.7.3 (2026-06-17)
+-------------------
+* Fixes
+  - Restore the grids_response_cc and grids_noresponse_cc APIs that were removed in 1.7.1
+
+
 v1.7.2 (2026-06-15)
 -------------------
 * Improvements

{gpu4pyscf-1.7.2/gpu4pyscf.egg-info → gpu4pyscf-1.7.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: gpu4pyscf
-Version: 1.7.2
+Version: 1.7.4
 Summary: GPU extensions for PySCF
 Author: PySCF developers
 License: Apache-2.0

{gpu4pyscf-1.7.2 → gpu4pyscf-1.7.4}/gpu4pyscf/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = '1.7.2'
+__version__ = '1.7.4'
 
 from . import _patch_pyscf
 

{gpu4pyscf-1.7.2 → gpu4pyscf-1.7.4}/gpu4pyscf/grad/rks.py

@@ -17,6 +17,7 @@
 
 '''Non-relativistic RKS analytical nuclear gradients'''
 from concurrent.futures import ThreadPoolExecutor
+import warnings
 import ctypes
 import numpy
 import cupy
@@ -800,11 +801,52 @@ def get_nlc_exc_full_response(ni, mol, grids, xc_code, dms, relativity=0, hermi=
 
     return exc1, 0
 
+# JCP 98, 5612 (1993); DOI:10.1063/1.464906
 def grids_response_cc(grids):
-    raise NotImplementedError("grids_response_cc() in GPU4PySCF is not used or tested anymore")
+    if not getattr(grids_response_cc, "_warned", False):
+        grids_response_cc._warned = True
+        warnings.warn("grids_response_cc() in GPU4PySCF is not used or tested anymore",
+                      DeprecationWarning)
+
+    # Notice: the returned grid order could be different from pyscf.grad.rks.grids_response_cc()!
+    mol = grids.mol
+
+    grid_to_atom_index_map = grids.atm_idx
+    atom_to_grid_index_map = [cupy.where(grid_to_atom_index_map == i_atom)[0] for i_atom in range(mol.natm)]
+    grid_to_atom_index_map = None
+
+    from gpu4pyscf.hessian.rks import get_dweight_dA # Avoid circular dependency
+
+    for i_atom in range(mol.natm):
+        i_g = atom_to_grid_index_map[i_atom]
+        fake_grids = type('FakeGrid', (object,), {})()
+        fake_grids.coords = grids.coords[i_g, :]
+        fake_grids.weights = grids.weights[i_g]
+        fake_grids.quadrature_weights = grids.quadrature_weights[i_g]
+        fake_grids.atm_idx = cupy.zeros(len(i_g), dtype = cupy.int32) + i_atom
+        fake_grids.atomic_radii = grids.atomic_radii
+        fake_grids.radii_adjust = grids.radii_adjust
+        fake_grids.becke_scheme = grids.becke_scheme
+        dw_dA_i = get_dweight_dA(mol, fake_grids)
+        yield fake_grids.coords, fake_grids.weights, dw_dA_i
 
 def grids_noresponse_cc(grids):
-    raise NotImplementedError("grids_noresponse_cc() in GPU4PySCF is not used or tested anymore")
+    if not getattr(grids_noresponse_cc, "_warned", False):
+        grids_noresponse_cc._warned = True
+        warnings.warn("grids_noresponse_cc() in GPU4PySCF is not used or tested anymore",
+                      DeprecationWarning)
+
+    # same as above but without the response, for nlc grids response routine
+    # Similarly, the returned grid order could be different from pyscf.grad.rks.grids_noresponse_cc()!
+    mol = grids.mol
+
+    grid_to_atom_index_map = grids.atm_idx
+    atom_to_grid_index_map = [cupy.where(grid_to_atom_index_map == i_atom)[0] for i_atom in range(mol.natm)]
+    grid_to_atom_index_map = None
+
+    for i_atom in range(mol.natm):
+        i_g = atom_to_grid_index_map[i_atom]
+        yield grids.coords[i_g, :], grids.weights[i_g]
 
 class Gradients(rhf_grad.Gradients):
     from gpu4pyscf.lib.utils import to_gpu, device

{gpu4pyscf-1.7.2 → gpu4pyscf-1.7.4}/gpu4pyscf/lib/cupy_helper.py

@@ -27,6 +27,7 @@ from gpu4pyscf.lib.memcpy import copy_array, p2p_transfer  #NOQA
 from gpu4pyscf.lib import multi_gpu
 from gpu4pyscf.lib.utils import load_library
 from gpu4pyscf.__config__ import num_devices, _p2p_access
+from gpu4pyscf import __config__
 
 LMAX_ON_GPU = 7
 DSOLVE_LINDEP = 1e-13
@@ -801,6 +802,13 @@ def _gen_x0(v, xs):
         x0 = x0[0]
     return x0
 
+_MAX_MEMORY = os.getenv('PYSCF_MAX_MEMORY')
+if _MAX_MEMORY is None and __config__.MAX_MEMORY != 4000:
+    # 4000 MB is the built-in default initialized in pyscf.__config__.
+    # If MAX_MEMORY has been explicitly configured (e.g. in ~/.pyscf_conf.py),
+    # use that value when the environment variable is not set.
+    _MAX_MEMORY = __config__.MAX_MEMORY
+
 def empty_mapped(shape, dtype=float, order='C'):
     '''(experimental)
     Returns a new, uninitialized NumPy array with the given shape and dtype.
@@ -812,6 +820,15 @@ def empty_mapped(shape, dtype=float, order='C'):
     size = int(np.prod(shape))
     nbytes = size * int(np.dtype(dtype).itemsize)
     assert nbytes >= 0, f"nbytes = {nbytes} is negative, type(nbytes) = {type(nbytes)}, please check if overflow happens"
+
+    if _MAX_MEMORY is not None:
+        mem_used = lib.current_memory()[0] # in MB
+        mem_limit = float(_MAX_MEMORY)
+        if nbytes*1e-6 + mem_used > mem_limit:
+            raise MemoryError(
+                f'Cannot allocate {nbytes} bytes of host memory '
+                f'(current: {mem_used:.1f} MB, limit: {mem_limit:.1f} MB).')
+
     mem = cupy.cuda.PinnedMemoryPointer(
         cupy.cuda.PinnedMemory(nbytes, cupy.cuda.runtime.hostAllocMapped), 0)
     out = np.ndarray(shape, dtype=dtype, buffer=mem, order=order)

{gpu4pyscf-1.7.2 → gpu4pyscf-1.7.4/gpu4pyscf.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: gpu4pyscf
-Version: 1.7.2
+Version: 1.7.4
 Summary: GPU extensions for PySCF
 Author: PySCF developers
 License: Apache-2.0