Fixes for saving the gemma tokenizer

scripts/bootstrap.sh

@@ -3,7 +3,7 @@
 #SIA_INSTALL_NO_WEB=1
 SIA_INSTALL_NO_NOTEBOOK=1
 #SIA_INSTALL_NO_CORE=1
-SIA_INSTALL_NO_ITB=1
+#SIA_INSTALL_NO_ITB=1
 SIA_INSTALL_NO_MISTRAL_INFER=1
 #SIA_INSTALL_NO_GEMMA_INFER=1
 #SIA_INSTALL_NO_GEMMA_TRAIN=1

scripts/install.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
 
 #SIA_INSTALL_NO_WEB=1
+#SIA_INSTALL_NO_LLAMA_CPP=1
 #SIA_INSTALL_NO_NOTEBOOK=1
 #SIA_INSTALL_NO_CORE=1
 #SIA_INSTALL_NO_ITB=1
@@ -20,6 +21,12 @@ if [ -z "${SIA_INSTALL_NO_WEB}" ]; then
     )
 fi
 
+if [ -z "${SIA_INSTALL_NO_LLAMA_CPP}" ]; then
+    echo "Installing venv for llama.cpp"
+    python3 -m venv /root/venvs/llama_cpp
+    /root/venvs/llama_cpp/bin/pip install /root/sia/modules/llama.cpp
+fi
+
 if [ -z "${SIA_INSTALL_NO_NOTEBOOK}" ]; then
     echo "Installing venv for running notebooks"
     python3 -m venv /root/venvs/notebook

scripts/train.sh

@@ -14,4 +14,6 @@ fi
 
 mkdir -p "$OUTPUT_DIR"
 
-gemma_train --model "google/gemma-3-12b-it" --output-dir "$OUTPUT_DIR"
+gemma_train --model "google/gemma-3-12b-it" --output-dir "$OUTPUT_DIR"
+cp "$OUTPUT_DIR"/tokenizer/* "$OUTPUT_DIR"/merged/
+llama-convert-hf-to-gguf --outfile "$OUTPUT_DIR/model.gguf" --outtype q8_0 "$OUTPUT_DIR/merged"

tools/gemma_train/pyproject.toml

@@ -8,13 +8,17 @@ version = "0.1.0"
 requires-python = ">=3.8"
 
 dependencies = [
-    "accelerate>=0.26.0",
+    "accelerate==1.4.0",
-    "bitsandbytes>=0.45.0",
+    "bitsandbytes==0.45.3",
-    "llama-cpp-scripts @ file:///root/sia/modules/llama.cpp",
+    "datasets==3.3.2",
+    "evaluate==0.4.3",
     "llm_engine_utils @ file:///root/sia/lib/llm_engine_utils",
-    "trl>=0.17.0",
+    "peft==0.14.0",
-    "peft>=0.15.0",
     "python-dotenv>=1.0.0",
+    "sentencepiece>=0.2.0",
+    "torch>=2.4.0",
+    "transformers>=4.51.3",
+    "trl==0.15.2",
 ]
 
 [project.scripts]

tools/gemma_train/src/gemma_train/__main__.py

@@ -13,11 +13,13 @@ def main():
     config = Config()
     train(config)
     merge(config)
-    os.system(f"cp -r {config.output_dir}/tokenizer/* {config.output_dir}/merged")
-    convert_to_gguf(config)
 
 def train(config: Config):
-    tokenizer = AutoTokenizer.from_pretrained(config.model, token=config.api_key)
+    tokenizer = AutoTokenizer.from_pretrained(
+        config.model,
+        token=config.api_key,
+        trust_remote_code=True,
+    )
     tokenizer.save_pretrained(config.output_dir/"tokenizer")
 
     bnb_config = BitsAndBytesConfig(
@@ -90,62 +92,6 @@ def merge(config: Config):
         safe_serialization=True
     )
 
-def convert_to_gguf(config: Config):
-    """Convert the merged model to GGUF format using llama.cpp's convert_hf_to_gguf script."""
-    print("Converting merged model to GGUF format...")
-    
-    # Add path to llama.cpp directory
-    sys.path.append("./llama.cpp")
-    
-    try:
-        # Import the necessary components from the conversion script
-        from convert_hf_to_gguf import ModelBase, gguf, ModelType
-        
-        # Set up paths
-        dir_model = config.output_dir / "merged"
-        fname_out = config.output_dir / "model.gguf"
-        output_type = gguf.LlamaFileType.MOSTLY_Q8_0  # Using Q8_0 quantization
-        
-        # Run the conversion with torch inference mode
-        with torch.inference_mode():
-            # Load hyperparameters
-            hparams = ModelBase.load_hparams(dir_model)
-            model_architecture = hparams["architectures"][0]
-            model_type = ModelType.TEXT
-            
-            print(f"Model architecture: {model_architecture}")
-            
-            try:
-                # Get the appropriate model class
-                model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
-                
-                # Create model instance
-                model_instance = model_class(
-                    dir_model, 
-                    output_type, 
-                    fname_out,
-                    is_big_endian=False, 
-                    use_temp_file=False, 
-                    eager=False
-                )
-                
-                # Export the model
-                print(f"Exporting model to GGUF format...")
-                model_instance.write()
-                print(f"Model successfully exported to {model_instance.fname_out}")
-                
-            except NotImplementedError:
-                print(f"Error: Model architecture {model_architecture} is not supported for GGUF conversion")
-                print("Skipping GGUF conversion")
-    
-    except ImportError as e:
-        print(f"Error importing conversion script: {e}")
-        print("Make sure llama.cpp is properly installed and accessible")
-        print("Skipping GGUF conversion")
-    except Exception as e:
-        print(f"Error during GGUF conversion: {e}")
-        print("Skipping GGUF conversion")
-
 def format_sia_example(example):
     return example['messages'].removeprefix("<bos>")