bfloat16 dtype 下 paddle.nn.functional.linear和 torch.nn.functional.linear 存在diff

[wtmlon]

bug描述 Describe the Bug

Description

• 注：本 issue 内需要用到的脚本都在这个PR，可直接下载

  在迁移 paddlenlp 模型时发现两端会在 nn.Linear后出现 diff，排查后发现paddle.matmul算子与 torch.matmul存在 diff

• 复现流程，首先在 paddle 虚拟环境下运行bf16_matmul.py，读取给定的 hs_0.npy(fp32)作为 x，w.npy (fp32)作为 W，进行(bf16) matmul 运算和 F.linear 运算，得出结果存为sdpa.npy(fp32)/sdpa1.npy(fp32), 该脚本代码如下

  
  import paddle
  import paddle.nn.functional as F
  import numpy as np

x_np = np.load('hs_0.npy') x = paddle.to_tensor(x_np, dtype='bfloat16') w_np = np.load('w.npy') w = paddle.to_tensor(w_np, dtype='bfloat16')

out = F.linear(x, w.T, bias=None) out1 = paddle.matmul(x, w, transpose_y=True) print(out) print(paddle.all((out-out1)==0)) np.save('sdpa.npy', out.astype('float32')) np.save('sdpa1.npy', out1.astype('float32'))

- 第二步，在 torch 虚拟环境下运行t_bf16_matmul.py，读取给定的 hs_0.npy(fp32)作为 x，w.npy (fp32)作为 W，进行(bf16) matmul 运算和 F.linear，得出结果存为t_sdpa.npy(fp32)/t_sdpa.npy(fp32), 该脚本代码如下
```python
import torch 
import torch.nn.functional as F
import numpy as np

x_np = np.load('hs_0.npy')
x = torch.tensor(x_np, dtype=torch.bfloat16).to("cuda:0")
w_np = np.load('w.npy')
w = torch.tensor(w_np, dtype=torch.bfloat16).to("cuda:0")

out1 = torch.matmul(x, w.transpose(0,1))
out = F.linear(x, w, bias=None)
print(out)
print(torch.all((out1-out) == 0))
np.save('t_sdpa.npy', out.cpu().detach().to(torch.float32))
np.save('t_sdpa1.npy', out1.cpu().detach().to(torch.float32))

最后运行 check.a.py脚本对两端结果做差

> python check.a.py
linear_diff:  0.00390625
matmul_diff:  0.0
paddle_linear_matmul_diff:  0.00390625
torch_linear_matmul_diff:  0.0

可以看到在 gpu 下，两端 matmul 可以对齐，但是两端的 F.linear 是有 diff 的

其他补充信息 Additional Supplementary Information

No response

[DesmonDay]

把torch的计算设备移到gpu上，可以对齐。