Update README.md

README.md

@@ -22,10 +22,6 @@ P.S. Recently we have received a lot of inquiries on accelerating customized mod
 
 ****
 
-## New Features (2023-06-20)
-- We now support cuda version of both 11.X and 12.X
-- lyraChatGLM has been further optimized, with faster model load speed from few minutes to less than 10s for non-int8 mode, and around 1 min for int8 mode!
-
 ## Model Card for lyraChatGLM
 
 lyraChatGLM is currently the **fastest ChatGLM-6B** available. To the best of our knowledge, it is the **first accelerated version of ChatGLM-6B**.
@@ -35,7 +31,10 @@ The inference speed of lyraChatGLM has achieved **300x** acceleration u
 Among its main features are:
 - weights: original ChatGLM-6B weights released by THUDM.
 - device: Nvidia GPU with Amperer architecture or Volta architecture (A100, A10, V100...).
-- batch_size: compiled with dynamic batch size, maximum depends on device. 
+- batch_size: compiled with dynamic batch size, maximum depends on device.
+New Features (2023-06-20)
+- We now support cuda version of both 11.X and 12.X
+- lyraChatGLM has been further optimized, with faster model load speed from few minutes to less than 10s for non-int8 mode, and around 1 min for int8 mode!
 
 ## Speed
 - orginal version(fixed batch infer): commit id 1d240ba