This guide provides instructions on how to download, set up, and run the Gemma 4 Vision (E2B) model using llama.cpp. It includes instructions for both text-only generation and multimodal (text + image) generation, along with testing examples for Windows and Linux.
You will need to download the following model files from HuggingFace:
| # Class that handles scaling operation. | |
| class Scaler: | |
| def __init__(self, scaler): | |
| self.scaler = scaler | |
| # For backward compatibilty with earlier version of scikit learn | |
| self.scaler.clip = False | |
| def transform_using_scaler(self, data_x): | |
| import glob | |
| from sklearn.preprocessing import MinMaxScaler | |
| def npy_header_offset(npy_path): | |
| """Gives the no of header bytes inside a numpy file.""" | |
| with open(str(npy_path), 'rb') as f: | |
| if f.read(6) != b'\x93NUMPY': | |
| raise ValueError('Invalid NPY file.') | |
| version_major, version_minor = f.read(2) | |
| if version_major == 1: |
| # we will utilize our class for creating generator. | |
| train_gen = CustomGenerator(train_data_descriptor, BATCH_SIZE, min_max_scaler_train) | |
| valid_gen = CustomGenerator(validation_data_descriptor, BATCH_SIZE, min_max_scaler_train) | |
| input_dim = train_gen.getitem(0)[0].shape[1] | |
| # The following two functions actually extract data on demand by using yield. | |
| def gen_data_train(): | |
| j = 0 | |
| while j < NO_OF_EPOCHS: |
| # Here is a class that makes it easy to load data in a batch. | |
| class CustomGenerator(): | |
| def __init__(self, data_desc, batch_size, scaler): | |
| self.data_desc = data_desc | |
| self.batch_size = batch_size | |
| self.scaler = scaler | |
| self.len = self.__len__() # an attribute for the length | |
| def __len__(self): |
| import glob | |
| from sklearn.preprocessing import MinMaxScaler | |
| training_files = glob.glob(os.path.join(train_dir, '*')) | |
| validation_files = glob.glob(os.path.join(validation_dir, '*')) | |
| min_max_scaler_train = MinMaxScaler() | |
| # constants declaration. Notice the constants are same for meaningful comparison. | |
| BATCH_SIZE = 32 | |
| NO_OF_EPOCHS = 3 |
| import glob | |
| from sklearn.preprocessing import MinMaxScaler | |
| training_files = glob.glob(os.path.join(train_dir, '*')) | |
| validation_files = glob.glob(os.path.join(validation_dir, '*')) | |
| # Declare constants like BATCH_SIZE and NO_OF_EPOCHS | |
| BATCH_SIZE = 32 | |
| NO_OF_EPOCHS = 3 |
| import tensorflow as tf | |
| from tensorflow.keras.layers import InputLayer, Dense, Input, Dropout,\ | |
| BatchNormalization, Flatten | |
| from tensorflow.keras import regularizers | |
| from tensorflow.keras.models import Sequential, Model, load_model | |
| from tensorflow.keras.optimizers import Adam | |
| from functools import partial | |
| def create_ae_model(num_hidden_layers=3, hidden_layer_neurons=64, | |
| lr=0.001, kernel_init='lecun_normal', optimizer='adam', |
| import os | |
| import numpy as np | |
| NO_OF_FILES = 1000 | |
| NO_OF_FEATURES = 1000 | |
| MAX_ROW_LIMIT_IN_SINGLE_FILE = 370 | |
| train_dir = os.path.join(os.getcwd(), 'train') | |
| os.makedirs(train_dir, exist_ok=True) |
