diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py
index 20aac4479c..aeb45c4e5d 100644
--- a/unsloth/models/gemma.py
+++ b/unsloth/models/gemma.py
@@ -83,6 +83,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
         t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0)
         inv_freq_expanded = inv_freq[None, :, None].float().expand(1, -1, 1).to("cuda")
         position_ids_expanded = t[:, None, :].float()
+        print(position_ids_expanded.to(torch.float64))
+        print(inv_freq_expanded.to(torch.float64))
         freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
         print(freqs.to(torch.int64))
         emb = torch.cat((freqs, freqs), dim=-1)
@@ -109,6 +111,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module):
         t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0)
         inv_freq_expanded = self.inv_freq[None, :, None].float().expand(1, -1, 1).to("cuda")
         position_ids_expanded = t[:, None, :].float()
+        print(position_ids_expanded.to(torch.float64))
+        print(inv_freq_expanded.to(torch.float64))
         freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
         print(freqs.to(torch.int64))
         emb = torch.cat((freqs, freqs), dim=-1)