import math
import torch

device = 'cuda'  # fails if device == 'cpu' (cf. #61854)

dim = 1
shape_src = 3, 5
shape_idx = 3, 7

index = torch.empty(shape_idx, dtype=torch.long, device=device).random_(0, shape_src[dim])

# succeeds
output = torch.zeros(shape_src, device=device)
output.scatter_(dim, index, 1.0, reduce='add')
print(output)

# desired usage
# output = torch.zeros(shape_src, device=device)
# values = torch.rand(shape_src, device=device)
# output.scatter_(dim, index, values, reduce='add')   # <-- failure
# print(output)

output = torch.zeros(shape_src, device=device)
values = torch.ones(shape_src, device=device)
chunks = math.ceil(index.shape[dim] / output.shape[dim])
for index_chunk in torch.chunk(index, chunks, dim):
    output.scatter_(dim, index_chunk, values, reduce='add')
print(output)