From 4eac5b59982e37dfdb8fe9c06e763235a09c64d3 Mon Sep 17 00:00:00 2001
From: Nicola Dall'Asen <fodark@pm.me>
Date: Wed, 13 Mar 2024 12:33:52 +0100
Subject: [PATCH] modify example in README to use the new device and dtype
 function

---
 README.md | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 6d24eaf..9bfaeeb 100644
--- a/README.md
+++ b/README.md
@@ -112,7 +112,7 @@ import torch
 from transformers import AutoModelForCausalLM
 
 from deepseek_vl.models import VLChatProcessor, MultiModalityCausalLM
-from deepseek_vl.utils.io import load_pil_images
+from deepseek_vl.utils.io import load_pil_images, get_device_and_dtype
 
 
 # specify the path to the model
@@ -121,7 +121,9 @@ vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
 tokenizer = vl_chat_processor.tokenizer
 
 vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
-vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
+
+device, dtype = get_device_and_dtype()
+vl_gpt = vl_gpt.to(dtype).to(device).eval()
 
 conversation = [
     {