Create README.md

README.md

@@ -0,0 +1,57 @@
+```
+import torch
+import transformers
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from PIL import Image
+import warnings
+
+# disable some warnings
+transformers.logging.set_verbosity_error()
+transformers.logging.disable_progress_bar()
+warnings.filterwarnings('ignore')
+
+# set device
+torch.set_default_device('cuda')  # or 'cpu'
+
+model_name = 'fne/dolphin-llava-72b'
+
+# create model
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    torch_dtype=torch.float16,
+    device_map='auto',
+    trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(
+    model_name,
+    trust_remote_code=True)
+
+# text prompt
+prompt = 'Describe this image in detail'
+
+messages = [
+    {"role": "user", "content": f'<image>\n{prompt}'}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+
+print(text)
+
+text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
+input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1], dtype=torch.long).unsqueeze(0)
+
+# image, sample images can be found in images folder
+image = Image.open('/path/to/image.png')
+image_tensor = model.process_images([image], model.config).to(dtype=model.dtype)
+
+# generate
+output_ids = model.generate(
+    input_ids,
+    images=image_tensor,
+    max_new_tokens=2048,
+    use_cache=True)[0]
+
+print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())
+```