Update app.py

app.py

@@ -1,20 +1,24 @@
-import os, subprocess, gradio as gr
+import os, subprocess, gradio as gr, shutil
 
 MODEL = "dolphin-2.9.3-llama3-8b.Q4_K_M.gguf"
 URL = "https://huggingface.co/TheBloke/dolphin-2.9.3-llama3-8b-GGUF/resolve/main/" + MODEL
 
-if not os.path.exists(MODEL):
-    subprocess.run(["wget", URL])
+def setup():
+    if not os.path.exists(MODEL):
+        subprocess.run(["wget", "-q", URL])
 
-if not os.path.exists("llama.cpp"):
-    subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp"])
-    subprocess.run(["make", "-C", "llama.cpp"])
+    if not os.path.exists("llama"):
+        subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp"])
+        subprocess.run(["cmake", "-S", "llama.cpp", "-B", "llama"])
+        subprocess.run(["cmake", "--build", "llama", "--config", "Release"])
+
+setup()
 
 def chat(prompt):
     p = subprocess.Popen(
-        ["./llama.cpp/main", "-m", MODEL, "-p", prompt, "-n", "200"],
+        ["./llama/main", "-m", MODEL, "-p", prompt, "-n", "200"],
         stdout=subprocess.PIPE
     )
-    return p.stdout.read().decode()
+    return p.stdout.read().decode(errors="ignore")
 
 gr.Interface(chat, gr.Textbox(label="Ask Dolphin"), gr.Textbox()).launch()