Create README.md

README.md

@@ -0,0 +1,31 @@
+# Quickstart
+
+```python
+import torch
+import requests
+from PIL import Image
+from transformers import AutoModelForCausalLM, LlamaTokenizer
+
+tokenizer = LlamaTokenizer.from_pretrained('lmsys/vicuna-7b-v1.5')
+model = AutoModelForCausalLM.from_pretrained(
+    'THUDM/cogvlm-base-224-hf',
+    torch_dtype=torch.bfloat16,
+    low_cpu_mem_usage=True,
+    trust_remote_code=True
+).to('cuda').eval()
+
+image = Image.open(requests.get('https://github.com/THUDM/CogVLM/blob/main/examples/1.png?raw=true', stream=True).raw).convert('RGB')
+inputs = model.build_conversation_input_ids(tokenizer, query='', images=[image])
+inputs = {
+    'input_ids': inputs['input_ids'].unsqueeze(0).to('cuda'),
+    'token_type_ids': inputs['token_type_ids'].unsqueeze(0).to('cuda'),
+    'attention_mask': inputs['attention_mask'].unsqueeze(0).to('cuda'),
+    'images': [[inputs['images'][0].to('cuda').to(torch.bfloat16)]],
+}
+gen_kwargs = {"max_length": 2048, "do_sample": False}
+
+with torch.no_grad():
+    outputs = model.generate(**inputs, **gen_kwargs)
+    outputs = outputs[:, inputs['input_ids'].shape[1]:]
+    print(tokenizer.decode(outputs[0]))
+```