Update README.md

README.md

@@ -12,6 +12,16 @@ In this case the tokenizer is the yi_tokenizer, loading it requires trust_remote
 
 Have some fun with this fellow. 
 
+It can eat a lot of vram, depends on settings to make it useable on two 24 gb vram gpus:
+
+Without fused attention, it's 27 gb vram, will need some if if yo do stuff. 
+
+![image.png](https://cdn-uploads.huggingface.co/production/uploads/63732ebbbd81fae2b3aaf3fb/1cbqKp55WhN4BQD337E-n.png)
+
+You can also let if have fused attention and just reduce the max_seq_length to something way smaller yet still useful
+
+![image.png](https://cdn-uploads.huggingface.co/production/uploads/63732ebbbd81fae2b3aaf3fb/JRi4sakPziGpmOFCBfcJS.png)
+
 
 
 License