Merge branch 'main' of https://github.com/unslothai/unsloth

unsloth/models/_utils.py

@@ -619,7 +619,11 @@ def patch_linear_scaling(
         f"from {model_filepath} import logger, "\
         f"{model_name.title()}Attention, {model_name.title()}Config"
 
-    function = inspect.getsource(attention_module.__init__)
+    try:
+        function = inspect.getsource(attention_module.__init__)
+    except:
+        # Most likely already patched!
+        return None, None
     where = function.find("def")
     function = function.split("\n")
     function = "\n".join(x[where:] for x in function)

unsloth/models/gemma.py

@@ -281,8 +281,10 @@ class FastGemmaModel(FastLlamaModel):
             scaled_rope_module = GemmaFixedLinearScalingRotaryEmbedding,
             attention_module   = GemmaAttention,
         )
-        exec(function, globals())
-        GemmaAttention.__init__      = eval(init_name)
+        if init_name is not None:
+            exec(function, globals())
+            GemmaAttention.__init__  = eval(init_name)
+        pass
         GemmaAttention      .forward = LlamaAttention_fast_forward
         GemmaSdpaAttention  .forward = LlamaAttention_fast_forward
         GemmaFlashAttention2.forward = LlamaAttention_fast_forward

unsloth/models/gemma2.py

@@ -436,8 +436,10 @@ class FastGemma2Model(FastLlamaModel):
             scaled_rope_module = GemmaFixedLinearScalingRotaryEmbedding,
             attention_module   = Gemma2Attention,
         )
-        exec(function, globals())
-        Gemma2Attention.__init__      = eval(init_name)
+        if init_name is not None:
+            exec(function, globals())
+            Gemma2Attention.__init__  = eval(init_name)
+        pass
         Gemma2Attention      .forward = Gemma2Attention_fast_forward
         Gemma2SdpaAttention  .forward = Gemma2Attention_fast_forward
         Gemma2FlashAttention2.forward = Gemma2Attention_fast_forward

unsloth/models/mistral.py

@@ -277,8 +277,10 @@ class FastMistralModel(FastLlamaModel):
             scaled_rope_module = LlamaLinearScalingRotaryEmbedding,
             attention_module   = MistralAttention,
         )
-        exec(function, globals())
-        MistralAttention.__init__      = eval(init_name)
+        if init_name is not None:
+            exec(function, globals())
+            MistralAttention.__init__  = eval(init_name)
+        pass
         MistralAttention      .forward = MistralAttention_fast_forward
         MistralSdpaAttention  .forward = MistralAttention_fast_forward
         MistralFlashAttention2.forward = MistralAttention_fast_forward

unsloth/models/qwen2.py

@@ -45,8 +45,10 @@ class FastQwen2Model(FastLlamaModel):
             scaled_rope_module = LlamaLinearScalingRotaryEmbedding,
             attention_module   = Qwen2Attention,
         )
-        exec(function, globals())
-        Qwen2Attention.__init__      = eval(init_name)
+        if init_name is not None:
+            exec(function, globals())
+            Qwen2Attention.__init__  = eval(init_name)
+        pass
         Qwen2Attention      .forward = LlamaAttention_fast_forward
         Qwen2SdpaAttention  .forward = LlamaAttention_fast_forward
         Qwen2FlashAttention2.forward = LlamaAttention_fast_forward