Merge pull request #109 from ibm-granite/clamp_min_issue

address issue with clamp_min on MPS

tsfm_public/models/tinytimemixer/modeling_tinytimemixer.py

@@ -982,7 +982,7 @@ def forward(
                 `(batch_size, 1, num_input_channels)`)
         """
         denominator = observed_indicator.sum(self.dim, keepdim=self.keepdim)
-        denominator = denominator.clamp_min(1.0)
+        denominator = denominator.clamp_min(torch.tensor(1, device=denominator.device))
         loc = (data * observed_indicator).sum(self.dim, keepdim=self.keepdim) / denominator
 
         variance = (((data - loc) * observed_indicator) ** 2).sum(self.dim, keepdim=self.keepdim) / denominator