conv2d计算结果与预期不符

[lh9171338]

bug描述 Describe the Bug

# paddlepaddle = 2.5.2
import paddle

dim = 1024
paddle.seed(1)
x = paddle.randn([1, dim, 3, 3])
kernel = paddle.zeros((dim, dim, 3, 3))
for i in range(dim):
    kernel[i, i, 1, 1] = 1
out = F.conv2d(x, kernel, padding=1)
diff = out - x
print(diff.abs().max())
# 预期值是0，相同逻辑用PyTorch实现是结果为0
# Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
#      0.00097394)

其他补充信息 Additional Supplementary Information

No response

[YuanRisheng]

麻烦提供一下正确的代码，这个跑不起来

[lh9171338]

# paddlepaddle = 2.5.2
import paddle
import paddle.nn.functional as F

dim = 1024
paddle.seed(1)
x = paddle.randn([1, dim, 3, 3])
kernel = paddle.zeros((dim, dim, 3, 3))
for i in range(dim):
    kernel[i, i, 1, 1] = 1
out = F.conv2d(x, kernel, padding=1)
diff = out - x
print(diff.abs().max())
# 预期值是0，相同逻辑用PyTorch实现是结果为0
# Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
#      0.00097394)

[LokeZhou]

# paddlepaddle = 2.5.2
import paddle
import paddle.nn.functional as F

dim = 1024
paddle.seed(1)
x = paddle.randn([1, dim, 3, 3])
kernel = paddle.zeros((dim, dim, 3, 3))
for i in range(dim):
    kernel[i, i, 1, 1] = 1
out = F.conv2d(x, kernel, padding=1)
diff = out - x
print(diff.abs().max())
# 预期值是0，相同逻辑用PyTorch实现是结果为0
# Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
#      0.00097394)

本地用在v100 cuda11.7,paddle develop版本跑的结果符合预期，可以用paddle2.6试试，或者paddle官网的nighty develop版本

[lh9171338]

我是A800跑的，cuda11.7，试了paddle2.6和nighty develop版本，结果都是0.00097394

[Ligoml]

@LokeZhou 辛苦再验证下 A800？

[LokeZhou]

可以试一下这设置这些flag
export FLAGS_use_cuda_managed_memory=true
export FLAGS_allocator_strategy=auto_growth
export FLAGS_embedding_deterministic=1
export FLAGS_cudnn_deterministic=1
export NVIDIA_TF32_OVERRIDE=0

[lh9171338]

设置这些flag之后输出结果是0了，感谢🙏