Open Acetylsalicylsaeure opened 3 months ago
sunqm
commented
3 months ago Please use gpu4pyscf.dft.RKS(mol) instead of mol.RKS(). The later creates a RKS object implemented in CPU. CPU module and GPU module are not compatible in this case.
There needs a check or a conversion on the ._scf attribute, when initializing hessian, gradients and other modules that rely on mf objects. @wxj6000
wxj6000
commented
3 months ago Please use
gpu4pyscf.dft.RKS(mol)instead ofmol.RKS(). The later creates a RKS object implemented in CPU. CPU module and GPU module are not compatible in this case.There needs a check or a conversion on the
._scfattribute, when initializing hessian, gradients and other modules that rely on mf objects. @wxj6000
This is one of the cases I didn't check the API compatibility carefully. I will work on it this coming week.
@Acetylsalicylsaeure Thank you for your feedback. Please use the solution by Qiming at this point. We are in the process of improving the compatibility of PySCF and GPU4PySCF.
Acetylsalicylsaeure
commented
3 months ago i was more worried about the occurence of pure pyscf code suddenly breaking due to gpu4pyscf being imported using gpu4pyscf.dft.rks.RKS(mol) still leads to the same error
In [1]: import numpy as np
...: import pyscf
...: from pyscf import lib, df
...: from gpu4pyscf.dft import rks
...: lib.num_threads(8)
...:
...: atom ='''
...: O 0.0000000000 -0.0000000000 0.1174000000
...: H -0.7570000000 -0.0000000000 -0.4696000000
...: H 0.7570000000 0.0000000000 -0.4696000000
...: '''
...:
...: mol = pyscf.M(atom=atom, basis='def2-tzvpp', verbose=0)
...:
In [2]: mf = rks.RKS(mol)
In [3]: mf.kernel()
Out[3]: -75.90464116157013
In [4]: hessobj = mf.Hessian()
In [5]: hess = hessobj.kernel()
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[5], line 1
----> 1 hess = hessobj.kernel()
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/hessian/rhf.py:524, in kernel(hessobj, mo_energy, mo_coeff, mo_occ, atmlst)
521 else:
522 hessobj.atmlst = atmlst
--> 524 de = hessobj.hess_elec(mo_energy, mo_coeff, mo_occ, atmlst=atmlst)
525 hessobj.de = de.get() + hessobj.hess_nuc(hessobj.mol, atmlst=atmlst)
526 mf = hessobj.base
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/hessian/rhf.py:59, in hess_elec(hessobj, mo_energy, mo_coeff, mo_occ, mo1, mo_e1, h1ao, atmlst, max_memory, verbose)
57 mo_occ = cupy.asarray(mo_occ)
58 mo_coeff = cupy.asarray(mo_coeff)
---> 59 de2 = hessobj.partial_hess_elec(mo_energy, mo_coeff, mo_occ, atmlst,
60 max_memory, log)
61 t1 = log.timer_debug1('hess elec', *t1)
62 if h1ao is None:
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/hessian/rks.py:61, in partial_hess_elec(hessobj, mo_energy, mo_coeff, mo_occ, atmlst, max_memory, verbose)
59 else:
60 hyb = mf._numint.hybrid_coeff(mf.xc, spin=mol.spin)
---> 61 de2, ej, ek = rhf_hess._partial_hess_ejk(hessobj, mo_energy, mo_coeff, mo_occ,
62 atmlst, max_memory, verbose,
63 abs(hyb) > 1e-10)
64 de2 += ej - hyb * ek # (A,B,dR_A,dR_B)
66 mem_now = lib.current_memory()[0]
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/hessian/rhf.py:181, in _partial_hess_ejk(hessobj, mo_energy, mo_coeff, mo_occ, atmlst, max_memory, verbose, with_k)
178 vk1 = vk1.reshape(3,3,nao,nao)
179 t1 = log.timer_debug1('contracting int2e_ipvip1 for atom %d'%ia, *t1)
--> 181 ej[i0,i0] += contract('xypq,pq->xy', vj1_diag[:,:,p0:p1], dm0[p0:p1])*2
182 ek[i0,i0] += contract('xypq,pq->xy', vk1_diag[:,:,p0:p1], dm0[p0:p1])
183 e1[i0,i0] -= contract('xypq,pq->xy', s1aa[:,:,p0:p1], dme0[p0:p1])*2
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/lib/cutensor.py:151, in contract(pattern, a, b, alpha, beta, out)
146 def contract(pattern, a, b, alpha=1.0, beta=0.0, out=None):
147 '''
148 a wrapper for general tensor contraction
149 pattern has to be a standard einsum notation
150 '''
--> 151 return contraction(pattern, a, b, alpha, beta, out=out)
File ~/miniconda3/envs/openmm/lib/python3.11/site-packages/gpu4pyscf/lib/cutensor.py:109, in contraction(pattern, a, b, alpha, beta, out, op_a, op_b, op_c, algo, jit_mode, compute_desc, ws_pref)
105 alpha = np.asarray(alpha)
106 beta = np.asarray(beta)
108 cutensor_backend.contract(cutensor._get_handle().ptr, plan.ptr,
--> 109 alpha.ctypes.data, a.data.ptr, b.data.ptr,
110 beta.ctypes.data, c.data.ptr, out.data.ptr,
111 ws.data.ptr, ws_size)
113 return out
AttributeError: 'memoryview' object has no attribute 'ptr'
In [6]:
however, calling .density_fit() on the RKS object fixes the error, is this expected behavior?
Acetylsalicylsaeure
commented
3 months ago Please use
gpu4pyscf.dft.RKS(mol)instead ofmol.RKS(). The later creates a RKS object implemented in CPU. CPU module and GPU module are not compatible in this case. There needs a check or a conversion on the._scfattribute, when initializing hessian, gradients and other modules that rely on mf objects. @wxj6000This is one of the cases I didn't check the API compatibility carefully. I will work on it this coming week.
@Acetylsalicylsaeure Thank you for your feedback. Please use the solution by Qiming at this point. We are in the process of improving the compatibility of PySCF and GPU4PySCF.
ah, i was too slow, thanks! i also experimented around a bit, works well with RHF
wxj6000
commented
3 months ago @Acetylsalicylsaeure Hessian calculation of regular SCF is not supported on GPU yet. It is only supported for density fitting scheme.
wxj6000
commented
3 months ago @Acetylsalicylsaeure This is a bug in GPU4PySCF RKS hessian module. PySCF hessian has been redefined. The issue has been fixed in the above PR. Thank you for your feedback.
i get an error when trying to calculate the hessian. it works fine with the usual code, but breaks when i import gpu4pyscf.
regular pySCF:
after importing gpu4pyscf:
i get the same error after just calling to_gpu() method