Batch first gru with 2 layers and bidirectional is allergic to batch size > 1

1

Why does this print 2 different values ?

import torch
import torch.nn as nn

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

class Test(nn.Module):
    def __init__(self):
        super().__init__()
        self.lin3 = nn.Linear(5, 400)
        self.rnn2 = nn.GRU(1, 100, num_layers=2, bidirectional=True, batch_first=True)

    def forward(self, discrete_code):
        c2 = self.lin3(discrete_code/5).reshape((4, discrete_code.shape[0], 100))
        x, _ = self.rnn2(torch.zeros(discrete_code.shape[0], 64, 1).to(device), c2)
        return x[:1, 0, 0]

model = Test().to(device)
candidats = torch.randn(10, 5).to(device)
print(model(candidats[:1]))  # e.g., tensor([0.1234])
print(model(candidats[:2]))  # e.g., tensor([0.1234]) EXPECTED SAME