From 9c565580fbcc6bf95ca9d775e83f0c863eab9f1e Mon Sep 17 00:00:00 2001
From: Daniel Han-Chen <danielhanchen@gmail.com>
Date: Sat, 24 Feb 2024 17:09:03 +1100
Subject: [PATCH] Update gemma.py

---
 unsloth/models/gemma.py | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py
index d896d5a080..5bfc885dc0 100644
--- a/unsloth/models/gemma.py
+++ b/unsloth/models/gemma.py
@@ -89,7 +89,7 @@ def GemmaDecoderLayer_fast_forward(
         hidden_states += residual
     else:
         residual = hidden_states
-        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = fast_rms_layernorm(self.input_layernorm, hidden_states)
         hidden_states, self_attn_weights, present_key_value = self.self_attn(
             hidden_states=hidden_states,
             causal_mask=causal_mask,
@@ -104,7 +104,7 @@ def GemmaDecoderLayer_fast_forward(
 
         # Fully Connected
         residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = fast_rms_layernorm(self.post_attention_layernorm, hidden_states)
         hidden_states = self.mlp(hidden_states)
         hidden_states = residual + hidden_states
     pass
@@ -350,7 +350,7 @@ def GemmaModel_fast_forward(
             all_self_attns += (layer_outputs[1],)
     pass
     
-    hidden_states = self.norm(hidden_states)
+    hidden_states = fast_rms_layernorm(self.norm, hidden_states)
 
     # add hidden states from the last decoder layer
     if output_hidden_states:
@@ -531,12 +531,12 @@ class FastGemmaModel(FastLlamaModel):
         pass
 
         print("Unsloth: Patching Gemma RMS Layernorm + 1")
-        # for name, module in model.named_modules():
-        #     if isinstance(module, GemmaRMSNorm):
-        #         module.weight += 1.0 # return output * (1 + self.weight)
-        #         if not hasattr(module, "variance_epsilon"):
-        #             module.variance_epsilon = module.eps # Gemma doesn't use variance_epsilon
-        # pass
+        for name, module in model.named_modules():
+            if isinstance(module, GemmaRMSNorm):
+                module.weight += 1.0 # return output * (1 + self.weight)
+                if not hasattr(module, "variance_epsilon"):
+                    module.variance_epsilon = module.eps # Gemma doesn't use variance_epsilon
+        pass
 
         # Clear deleted GPU items
         import gc