RuntimeError: dst.nbytes() >= (dst.storage_offset() * dst.element_size()) INTERNAL ASSERT FAILED at “/Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/native/mps/operations/Copy.mm”:130, please report a bug to PyTorch.
Dear Developers,
I got an error like this, could you give me some instructions on why this happens and how I can solve it?
Best,
Tianming
In my experience the MPS backend is still relatively new and has some bugs that need to be worked out. Could you send a small code snippet that reproduces this issue on your end?
Sure, here is the code:
def gaussian(Sj1,Slj1,Sj2,Slj2):
sigma1 = torch.tensor(0.5)
sigma2 = torch.tensor(0.5)
Kb = torch.tensor(0.001987204)
T = torch.tensor(300)
h0 = torch.tensor(0.0001)
g = torch.tensor(0)
C1 = torch.tensor(0)
C2 = torch.tensor(0)
for i in range((torch.where(Slj2 == Sj2)[0][0] - 5),(torch.where(Slj2 == Sj2)[0][0] + 6)):
if i < 0:
C2 = i + 1000
elif i > 999:
C2 = i - 1000
else:
C2 = i
for j in range((torch.where(Slj1 == Sj1)[0][0] - 5),(torch.where(Slj2 == Sj2)[0][0] + 6)):
if j < 0:
C1 = j + 1000
elif j > 999:
C1 = j -1000
else:
C1 = j
g = g + count[C2,C1] * h0 * torch.exp( (-(Sj1 - Slj1[C1]) ** 2 / (2 * sigma1 ** 2)) + (-(Sj2 - Slj2[C2]) ** 2 / (2 * sigma2 ** 2)) )
F[torch.where(Slj2 == Sj2)[0][0]][torch.where(Slj1 == Sj1)[0][0]] = torch.exp(-g / (Kb * T))
#print(F[Slj1.index(Sj1)][Slj2.index(Sj2)])
return F[torch.where(Slj2 == Sj2)[0][0]][torch.where(Slj1 == Sj1)[0][0]]
Seems to me like you might not be using torch.where correctly. It needs 2 more parameters, x and y. See the docs here:
oh, I really thought torch.where is the same as np.where
Corrected. I see it can be used as torch.nonzero(condition, as_tuple=True).
This should probably be raised on the Github page with a minimalist example(i.e. remove as much code as you can to still reproduce the error).
emmm, I still have the same issue after I change torch.where to torch.nonzero 
Yes, I meant I stood corrected.