fix: add revision parameter to FastBaseModel in vision.py

Propagate revision parameter to all from_pretrained calls in vision.py
to ensure consistent version pinning for vision models.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
(cherry picked from commit c5aa4ec927)

unsloth/chat_templates.py

@@ -1667,14 +1667,18 @@ 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

@@ -2331,6 +2331,7 @@ 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
@@ -2424,6 +2425,7 @@ class FastLlamaModel:
                 max_position_embeddings = max_position_embeddings,
                 trust_remote_code = trust_remote_code,
                 attn_implementation = preferred_attn_impl,
+                revision = revision,
                 **kwargs,
             )
         elif not fast_inference:
@@ -2436,6 +2438,7 @@ class FastLlamaModel:
                 max_position_embeddings = max_position_embeddings,
                 trust_remote_code = trust_remote_code,
                 attn_implementation = preferred_attn_impl,
+                revision = revision,
                 **kwargs,
             )
             model.fast_generate = make_fast_generate_wrapper(model.generate)
@@ -2468,6 +2471,7 @@ class FastLlamaModel:
                 disable_log_stats = disable_log_stats,
                 use_bitsandbytes = load_in_4bit,
                 unsloth_vllm_standby = unsloth_vllm_standby,
+                revision = revision,
                 fp8_mode = fp8_mode,
             )
             for allowed_arg in allowed_args:
@@ -2505,6 +2509,7 @@ class FastLlamaModel:
             token = token,
             trust_remote_code = trust_remote_code,
             fix_tokenizer = fix_tokenizer,
+            revision = revision,
         )
 
         model, tokenizer = patch_tokenizer(model, tokenizer)

unsloth/models/loader.py

@@ -553,6 +553,7 @@ class FastLanguageModel(FastLlamaModel):
             model_config = AutoConfig.from_pretrained(
                 model_name,
                 token = token,
+                revision = revision,
                 trust_remote_code = trust_remote_code,
             )
 
@@ -1304,6 +1305,7 @@ class FastModel(FastBaseModel):
             model_config = AutoConfig.from_pretrained(
                 model_name,
                 token = token,
+                revision = revision,
                 trust_remote_code = trust_remote_code,
             )
 

unsloth/models/vision.py

@@ -421,6 +421,7 @@ class FastBaseModel:
         auto_config = None,
         offload_embedding = False,
         float32_mixed_precision = None,  # Forces float32 mixed precision
+        revision = None,
         # vLLM parameters
         fast_inference = False,
         gpu_memory_utilization = 0.5,
@@ -720,6 +721,7 @@ class FastBaseModel:
                     model_name,
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    revision = revision,
                 )
             if hasattr(auto_config, "quantization_config"):
                 from transformers.quantizers.auto import (
@@ -776,12 +778,12 @@ class FastBaseModel:
                 model_name,
                 token = token,
                 trust_remote_code = trust_remote_code,
+                revision = revision,
             )
         setattr(auto_config, "_attn_implementation", config_attn_impl)
         if hasattr(auto_config, "attn_implementation"):
             setattr(auto_config, "attn_implementation", config_attn_impl)
         model_config = auto_config
-
         verify_fp8_support_if_applicable(model_config)
 
         raise_handler = RaiseUninitialized()
@@ -796,6 +798,7 @@ class FastBaseModel:
                 # quantization_config   = bnb_config,
                 token = token,
                 trust_remote_code = trust_remote_code,
+                revision = revision,
                 # attn_implementation   = attn_implementation,
                 **kwargs,
             )

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