The training process loss is nan

[Xuagent]

Model：
class MobileNetV3L_64(nn.Module):
def init(self):
super(MobileNetV3L_64, self).init()
self.backbone = timm.create_model('mobilenetv3_large_100', pretrained=True, exportable=True)
self.backbone.classifier = Identity()
self.fc = nn.Linear(1280, 64)
self.l2_norm = L2Norm()

def forward(self, x):
    x = self.backbone(x)
    print(self.fc.bias)
    x = self.fc(x)
    x = self.l2_norm(x)
    return x

Freezing all params...
Unfreezing fc
Unfreezing l2_norm
Total parameters: 4284016
Trainable: 81984
Non-trainable: 4202032
Loss metric: semihard triplet loss.
Overall progrress: 0%| | 0/200 [00:00<?, ?it/sP
arameter containing: | 0/1157 [00:00<?, ?it/s]
tensor([-0.0278, 0.0100, -0.0074, 0.0099, -0.0233, 0.0184, 0.0254, 0.0190,
-0.0035, -0.0131, -0.0202, 0.0249, 0.0030, -0.0152, 0.0108, -0.0017,
0.0087, 0.0180, -0.0020, 0.0107, 0.0183, 0.0091, 0.0024, -0.0217,
0.0095, 0.0122, -0.0010, -0.0135, 0.0237, 0.0144, 0.0194, 0.0059,
-0.0019, -0.0021, 0.0274, -0.0133, 0.0193, -0.0204, -0.0190, 0.0040,
-0.0178, 0.0049, 0.0126, -0.0026, -0.0035, 0.0175, 0.0258, -0.0009,
0.0181, 0.0096, -0.0056, -0.0118, 0.0132, -0.0062, 0.0272, 0.0249,
-0.0076, -0.0042, 0.0186, 0.0279, 0.0120, 0.0230, -0.0012, 0.0220],
device='cuda:0', requires_grad=True)
P
arameter containing: | 1/1157 [00:02<57:10, 2.97s/it]
tensor([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan],
device='cuda:0', requires_grad=True)

The bias of fc is all nan the second time, resulting in loss being nan.
How to modify to solve this problem?
Thanks