still fixing

rag/model_loader.py

@@ -180,11 +180,17 @@ def load_enhanced_model_m2max() -> Tuple[AutoModelForCausalLM, AutoTokenizer]:
         cache_dir = os.getenv("HF_HOME") or os.getenv("TRANSFORMERS_CACHE") or "/workspace/.cache/huggingface" if os.getenv("WORKSPACE") else ".cache/huggingface"
         
         # For LoRA models, try loading tokenizer from LoRA directory first, then base model
+        # Use slow tokenizer (use_fast=False) which requires sentencepiece for Llama/Mistral models
         tokenizer = None
         if model_path and (model_path / "tokenizer.json").exists():
             try:
                 logger.info(f"📝 Loading tokenizer from LoRA directory: {model_path}")
-                tokenizer = AutoTokenizer.from_pretrained(str(model_path), cache_dir=cache_dir, trust_remote_code=True)
+                tokenizer = AutoTokenizer.from_pretrained(
+                    str(model_path), 
+                    cache_dir=cache_dir, 
+                    trust_remote_code=True,
+                    use_fast=False  # Use slow tokenizer with sentencepiece
+                )
             except Exception as e:
                 logger.warning(f"⚠️  Could not load tokenizer from LoRA dir: {e}, using base model")
         
@@ -193,7 +199,8 @@ def load_enhanced_model_m2max() -> Tuple[AutoModelForCausalLM, AutoTokenizer]:
                 base_model_name,
                 cache_dir=cache_dir,
                 padding_side='left',  # Required for decoder-only models
-                trust_remote_code=True
+                trust_remote_code=True,
+                use_fast=False  # Use slow tokenizer with sentencepiece
             )
         
         if tokenizer.pad_token is None:
@@ -297,7 +304,8 @@ def load_enhanced_model_m2max() -> Tuple[AutoModelForCausalLM, AutoTokenizer]:
             base_model_name,
             cache_dir=cache_dir,
             padding_side='left',
-            trust_remote_code=True
+            trust_remote_code=True,
+            use_fast=False  # Use slow tokenizer with sentencepiece
         )
         
         if tokenizer.pad_token is None:
@@ -380,7 +388,8 @@ def load_enhanced_model_m2max() -> Tuple[AutoModelForCausalLM, AutoTokenizer]:
                 base_model_name,
                 cache_dir=cache_dir,
                 padding_side='left',
-                trust_remote_code=True
+                trust_remote_code=True,
+                use_fast=False  # Use slow tokenizer with sentencepiece
             )
             
             if tokenizer.pad_token is None: