use int8 model

app.py

@@ -1,3 +1,19 @@
+with open('/usr/local/lib/python3.8/dist-packages/torch/nn/parameter.py', 'r') as f:
+    text = f.read()
+
+if 'if data.dtype == torch.int8' not in text:
+    text = text.replace(
+        '            return torch.Tensor._make_subclass(cls, data, requires_grad)', 
+        '            if data.dtype == torch.int8:\n' \
+        '                requires_grad = False\n' \
+        '            return torch.Tensor._make_subclass(cls, data, requires_grad)'
+    )
+
+    with open('/usr/local/lib/python3.8/dist-packages/torch/nn/parameter.py', 'w') as f:
+        f.write(text)
+
+
+
 # Credit:
 # https://huggingface.co/spaces/ljsabc/Fujisaki/blob/main/app.py
 
@@ -8,7 +24,7 @@ from peft import PeftModel
 from transformers import AutoTokenizer, GenerationConfig, AutoModel
 
 
-model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True, revision="4de8efe").float()
+model = AutoModel.from_pretrained("KumaTea/twitter-int8", trust_remote_code=True).float()
 tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True, revision="4de8efe")
 peft_path = 'KumaTea/twitter'
 model = PeftModel.from_pretrained(
@@ -90,7 +106,7 @@ def evaluate_stream(msg, history, temperature, top_p):
 
 
 title = """<h1 align="center">KumaGLM</h1>
-<h3 align='center'>这是一个 AI Kuma，你可以与她聊天，或者直接在文本框按下Enter</h3>
+<h3 align='center'>这是一个 AI Kuma，你可以与他聊天，或者直接在文本框按下Enter</h3>
 <p align='center'>采样范围 2020/06/13 - 2023/04/15</p>"""
 footer =  """<p align='center'>
 本项目基于