Upload 2 files

app.py

@@ -3,7 +3,7 @@ import os
 import gradio as gr
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, AutoModelForCausalLM
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 import timm
 from torchvision import transforms
 from llama_cpp import Llama
@@ -72,9 +72,15 @@ image_encoder.eval() # Set to evaluation mode
 
 base_model_name="microsoft/Phi-3-mini-4k-instruct"
 device = "cpu"
-
-base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype=torch.float32, device_map={"": device})
-
+bnb_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_use_double_quant=True,
+        bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=torch.bfloat16       
+        )
+#base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype=torch.float32, device_map={"": device})
+base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype=torch.bfloat16, trust_remote_code=True, # Important for some Phi-3 variants                                                          
+                                                         quantization_config=bnb_config, device_map={"": device})
 
 # Load and merge
 model = PeftModel.from_pretrained(base_model, peft_model_path, offload_dir='./offload')

requirements.txt

@@ -5,4 +5,5 @@ timm
 Pillow 
 transformers
 llama-cpp-python
-peft
+peft
+bitsandbytes