diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py
index c23f9171f8..ba7aac7d31 100644
--- a/unsloth/models/gemma.py
+++ b/unsloth/models/gemma.py
@@ -156,8 +156,9 @@ def GemmaModel_fast_forward_inference(
     hidden_states = self.embed_tokens(input_ids)
     # 3072**0.5 = 55.5000 in bfloat16, whilst 55.4256 in float32
     # 2048**0.5 = 45.2500 in bfloat16, whilst 45.2548 in float32
-    inputs_embeds *= torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype)
-
+    # inputs_embeds *= torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype)
+    inputs_embeds *= math_sqrt(self.config.hidden_size)
+    
     next_decoder_cache = []
     for idx, decoder_layer in enumerate(self.layers):
         # Self Attention
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index d7743216fb..4b9161a1a5 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -523,7 +523,8 @@ def LlamaModel_fast_forward(
         # inputs_embeds *= math_sqrt(self.config.hidden_size)
         # Ie 3072**0.5 = 55.5000 in bfloat16, whilst 55.4256 in float32
         # &  2048**0.5 = 45.2500 in bfloat16, whilst 45.2548 in float32
-        inputs_embeds *= torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype)
+        # inputs_embeds *= torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype)
+        inputs_embeds *= math_sqrt(self.config.hidden_size)
         if inputs_requires_grad: inputs_embeds.requires_grad_(True)
     pass