From a8e6364e66453ed2f2ac06ef1811e39d8c2a59ab Mon Sep 17 00:00:00 2001
From: DoubleMathew <mmathew23@gmail.com>
Date: Thu, 9 Oct 2025 10:39:50 -0500
Subject: [PATCH] Handle transformers rename from PretrainedConfig to
 PreTrainedConfig

---
 unsloth_zoo/compiler.py                  | 2 +-
 unsloth_zoo/empty_model.py               | 7 ++++++-
 unsloth_zoo/hf_utils.py                  | 7 ++++++-
 unsloth_zoo/patching_utils.py            | 7 ++++++-
 unsloth_zoo/saving_utils.py              | 2 +-
 unsloth_zoo/temporary_patches/gpt_oss.py | 8 +++++++-
 6 files changed, 27 insertions(+), 6 deletions(-)

diff --git a/unsloth_zoo/compiler.py b/unsloth_zoo/compiler.py
index 3125db73b..e9114445b 100644
--- a/unsloth_zoo/compiler.py
+++ b/unsloth_zoo/compiler.py
@@ -2148,7 +2148,7 @@ def replaced_tqdm(*args, **kwargs):
     # Get all functions as well
     functions = [x for x in functions if x not in torch_modules or not compile_torch_modules or not compile_custom_modules]
 
-    # Get all PretrainedModel classes
+    # Get all PreTrainedModel classes
     pretrained_modules = re.findall(r"class ([^\s]{1,})\(.+?PreTrainedModel\)", full_source)
 
     # Remove if no forward function
diff --git a/unsloth_zoo/empty_model.py b/unsloth_zoo/empty_model.py
index d73cf05eb..36cfce8b1 100644
--- a/unsloth_zoo/empty_model.py
+++ b/unsloth_zoo/empty_model.py
@@ -47,7 +47,12 @@ def compare_dicts(orig_dict, new_dict, prefix=""):
             print(f"Dict key {key_path} type mismatch: original {type(orig_val)} != new model {type(new_val)}")
 
 def compare_attributes(original_model, new_model):
-    from transformers.configuration_utils import PretrainedConfig
+    try:
+        from transformers.configuration_utils import PreTrainedConfig
+        PretrainedConfig = PreTrainedConfig
+    except:
+        from transformers.configuration_utils import PretrainedConfig
+
     print("=== ATTRIBUTE COMPARISON REPORT ===")
     missing_attrs = []
     type_mismatches = []
diff --git a/unsloth_zoo/hf_utils.py b/unsloth_zoo/hf_utils.py
index ee7763a31..5f0477fca 100644
--- a/unsloth_zoo/hf_utils.py
+++ b/unsloth_zoo/hf_utils.py
@@ -21,7 +21,12 @@
 import json
 import re
 from .log import logger
-from transformers import PretrainedConfig
+try:
+    from transformers import PreTrainedConfig
+    PretrainedConfig = PreTrainedConfig
+except:
+    from transformers import PretrainedConfig
+
 HAS_TORCH_DTYPE = "torch_dtype" in PretrainedConfig.__doc__
 
 __all__ = [
diff --git a/unsloth_zoo/patching_utils.py b/unsloth_zoo/patching_utils.py
index 0f1288c68..fd9c9ef05 100644
--- a/unsloth_zoo/patching_utils.py
+++ b/unsloth_zoo/patching_utils.py
@@ -242,7 +242,12 @@ def verify_and_set_device(module,):
 
 def patch_to_dict():
     from functools import wraps
-    from transformers.configuration_utils import PretrainedConfig
+    try:
+        from transformers.configuration_utils import PreTrainedConfig
+        PretrainedConfig = PreTrainedConfig
+    except:
+        from transformers.configuration_utils import PretrainedConfig
+
     from .hf_utils import _normalize_dict_dtypes
     original_to_dict = PretrainedConfig.to_dict
     original_to_diff_dict = PretrainedConfig.to_diff_dict
diff --git a/unsloth_zoo/saving_utils.py b/unsloth_zoo/saving_utils.py
index eb674081d..bcf8b7e8b 100644
--- a/unsloth_zoo/saving_utils.py
+++ b/unsloth_zoo/saving_utils.py
@@ -1780,7 +1780,7 @@ def detect_keys_format(keys_to_check, forward_mapping):
 def _convert_lora_keys_to_safetensor_format(
     lora_weights,        # Global dict of LoraStats objects
     safetensor_keys,     # List of keys from the CURRENT shard
-    model_class_name="PretrainedModel" # The actual model instance (e.g. Qwen2VLForConditionalGeneration)
+    model_class_name="PreTrainedModel" # The actual model instance (e.g. Qwen2VLForConditionalGeneration)
 ):
     import re
 
diff --git a/unsloth_zoo/temporary_patches/gpt_oss.py b/unsloth_zoo/temporary_patches/gpt_oss.py
index 0a369a7e9..06901f0d6 100644
--- a/unsloth_zoo/temporary_patches/gpt_oss.py
+++ b/unsloth_zoo/temporary_patches/gpt_oss.py
@@ -1414,7 +1414,13 @@ def encode_conversations_with_harmony(
 # RuntimeError: Unsloth: Failed to load model. Both AutoConfig and PeftConfig loading failed.
 # AutoConfig error: 'GptOssConfig' object has no attribute 'max_position_embeddings'
 try:
-    from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+    from transformers.configuration_utils import layer_type_validation
+    try:
+        from transformers.configuration_utils import PreTrainedConfig
+        PretrainedConfig = PreTrainedConfig
+    except:
+        from transformers.configuration_utils import PretrainedConfig
+
     from transformers.modeling_rope_utils import rope_config_validation
 
     class Old_GptOssConfig(PretrainedConfig):