Update README.md

README.md

@@ -46,6 +46,18 @@ reference-data-model:
 ```
 
 
+## 
+
+
+```py
+
+#Train activation
+#https://github.com/tomaarsen/attention_sinks
+
+pip install attention_sinks
+```
+
+
 ## Version
 ```py
 import torch, transformers,torchvision
@@ -67,6 +79,9 @@ from transformers import (
     GenerationConfig,
     TextIteratorStreamer,
 )
+
+from attention_sinks import AutoModelForCausalLM
+
 import torch
 
 # model_id = 'Open-Orca/Mistral-7B-OpenOrca'
@@ -78,6 +93,9 @@ model = AutoModelForCausalLM.from_pretrained(model_id,
                                              torch_dtype=torch.bfloat16,
                                              load_in_4bit=True,
                                              low_cpu_mem_usage= True,
+
+                                             attention_sink_size=4,
+                                             attention_sink_window_size=1024, #512, # <- Low for the sake of faster generation
                                              )
 
 max_length=2048