From 82e6a93ef25c7d3da50b4a4b717c44736d3bfc42 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 7 Jul 2024 09:33:01 -0700
Subject: [PATCH] Fix exec, eval (#743)

* Update gemma2.py

* Update llama.py

* Update llama.py

* Update gemma2.py

* init

* Update gemma2.py

* Update gemma2.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update gemma2.py

* Update gemma2.py

* Update gemma2.py

* All RoPE Scaling support

* cleanup

* Update llama.py

* Update llama.py

* Update _utils.py

* Update _utils.py

* exec

* exec

* Attention_Module

* attention_module

* imports

* exec

* Update llama.py

* Update llama.py

* boolean mask

* revert masking

* Update llama.py

* Update save.py

* Update llama.py

* Update gemma2.py

* Update gemma2.py

* Update gemma2.py

* Update utils.py

* retry

* Update gemma2.py

* Update gemma2.py

* Update gemma2.py

* Update _utils.py

* Update _utils.py

* Update gemma2.py

* Update chat_templates.py

* Gemma 2 Ollama support

* Update llama.py

* Update llama.py

* error handling

* Update _utils.py

* Update _utils.py

* Stats for debugging

* Update _utils.py

* Update _utils.py

* Debugging

* Update tokenizer_utils.py

* Update _utils.py

* Update cross_entropy_loss.py

* Update cross_entropy_loss.py

* Update cross_entropy_loss.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Update rms_layernorm.py

* Check exec, eval
---
 unsloth/models/_utils.py  | 6 +++++-
 unsloth/models/gemma.py   | 6 ++++--
 unsloth/models/gemma2.py  | 6 ++++--
 unsloth/models/mistral.py | 6 ++++--
 unsloth/models/qwen2.py   | 6 ++++--
 5 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index e0df218a93..725e9b3ff3 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -618,7 +618,11 @@ 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)
diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py
index 4d3db8d393..c47e65e87b 100644
--- a/unsloth/models/gemma.py
+++ b/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
diff --git a/unsloth/models/gemma2.py b/unsloth/models/gemma2.py
index 4a1420fb42..fda78534d8 100644
--- a/unsloth/models/gemma2.py
+++ b/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
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index 28f664ca2d..6eb3fccfab 100644
--- a/unsloth/models/mistral.py
+++ b/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
diff --git a/unsloth/models/qwen2.py b/unsloth/models/qwen2.py
index dcd05af60e..82de1951b8 100644
--- a/unsloth/models/qwen2.py
+++ b/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