fix: add revision parameter support and escape quotes in chat templates

- Fix #3544: Add revision parameter to AutoConfig, AutoModelForCausalLM,
  AutoModelForSequenceClassification, and load_correct_tokenizer calls
  in FastLlamaModel.from_pretrained. This enables loading specific model
  revisions/branches from HuggingFace Hub.

- Fix #3667: Escape single quotes in system messages before substituting
  into Jinja2 templates. This prevents TemplateSyntaxError when system
  messages contain apostrophes (e.g., "user's" in Vicuna templates).

Signed-off-by: majiayu000 <1835304752@qq.com>

unsloth/chat_templates.py

@@ -2108,14 +2108,18 @@ def _change_system_message(template: str, type_chat_template: str, system_messag
         if has_placeholder:
             if system_message is None:
                 raise ValueError("Unsloth: You need to provide a system message for custom templates.")
-            new_template = re.sub(system_message_pattern, system_message, template)
+            # Escape single quotes to prevent Jinja2 template syntax errors
+            escaped_message = system_message.replace("'", "\\'")
+            new_template = re.sub(system_message_pattern, escaped_message, template)
             return new_template, system_message
 
         return template, system_message
 
     # For predefined templates with default system message
     message_to_use = system_message if system_message is not None else default_system_message
-    new_template = re.sub(system_message_pattern, message_to_use, template)
+    # Escape single quotes to prevent Jinja2 template syntax errors
+    escaped_message = message_to_use.replace("'", "\\'")
+    new_template = re.sub(system_message_pattern, escaped_message, template)
 
     return new_template, message_to_use
 

unsloth/models/llama.py

@@ -2223,6 +2223,7 @@ class FastLlamaModel:
             model_name,
             token = token,
             attn_implementation = "sdpa",
+            revision = revision,
         )
         model_config.model_name = model_name
         model_max_seq_length = model_config.max_position_embeddings
@@ -2312,6 +2313,7 @@ class FastLlamaModel:
                 max_position_embeddings = max_position_embeddings,
                 trust_remote_code = trust_remote_code,
                 attn_implementation = "eager",
+                revision = revision,
                 **kwargs,
             )
         elif not fast_inference:
@@ -2324,6 +2326,7 @@ class FastLlamaModel:
                 max_position_embeddings = max_position_embeddings,
                 trust_remote_code = trust_remote_code,
                 attn_implementation = "eager",
+                revision = revision,
                 **kwargs,
             )
             model.fast_generate = model.generate
@@ -2381,6 +2384,7 @@ class FastLlamaModel:
             token = token,
             trust_remote_code = trust_remote_code,
             fix_tokenizer = fix_tokenizer,
+            revision = revision,
         )
 
         model, tokenizer = patch_tokenizer(model, tokenizer)

unsloth/tokenizer_utils.py

@@ -503,6 +503,7 @@ def _load_correct_tokenizer(
     trust_remote_code = False,
     cache_dir = "huggingface_tokenizers_cache",
     fix_tokenizer = True,
+    revision = None,
 ):
     if IS_COLAB_ENVIRONMENT:
         cache_dir = cache_dir
@@ -528,6 +529,7 @@ def _load_correct_tokenizer(
             legacy = False,
             from_slow = True,
             cache_dir = cache_dir,
+            revision = revision,
         )
     except:
         slow_tokenizer = None
@@ -546,6 +548,7 @@ def _load_correct_tokenizer(
         token = token,
         trust_remote_code = trust_remote_code,
         cache_dir = cache_dir,
+        revision = revision,
     )
 
     if not fix_tokenizer or tokenizer_name in IGNORED_TOKENIZER_NAMES:
@@ -587,6 +590,7 @@ def load_correct_tokenizer(
     trust_remote_code = False,
     cache_dir = "huggingface_tokenizers_cache",
     fix_tokenizer = True,
+    revision = None,
 ):
     tokenizer = _load_correct_tokenizer(
         tokenizer_name = tokenizer_name,
@@ -596,6 +600,7 @@ def load_correct_tokenizer(
         trust_remote_code = trust_remote_code,
         cache_dir = cache_dir,
         fix_tokenizer = fix_tokenizer,
+        revision = revision,
     )
 
     ### 1. Fixup tokenizer's chat_template