pytorch strangeness

pytorch_strangeness.py

import torch
import torch.nn as nn

torch.manual_seed(42)
x = torch.randn(2, 768)

# matrix multiply "ignores" the second row when calculating the first row
w = torch.randn(768, 768)
z1 = x[0] @ w
z2 = (x @ w)[0]
print((z1-z2).abs().max().item()) # prints 0 (should be 0, OK)

# linear does not!
m = nn.Linear(768, 768, bias=False)
with torch.no_grad():
    m.weight.copy_(w.T)
q1 = m(x[0])
q2 = m(x)[0]
print((q1-q2).abs().max().item()) # prints ~2e-5 ( should be 0?!)

# and z1 != q1
print((z1-q1).abs().max().item()) # prints ~9e-5 (should be 0?!)

I looked at torch source code and found this

inline Tensor linear(
    const Tensor& input,
    const Tensor& weight,
    const Tensor& bias = {}) {
  if (input.dim() == 2 && bias.defined()) {
    // fused op is marginally faster
    return torch::addmm(bias, input, weight.t());
  } else {
    auto output = input.matmul(weight.t());
    if (bias.defined()) {
      output += bias;
    }
    return output;
  }
}

could it be caused by the difference in which fused operation is applied for batched inputs?