Update handler.py

handler.py

@@ -1,18 +1,18 @@
-import requests
-
-class EndpointHandler:
-    def __init__(self, path=""):
-        self.api_url = "http://localhost:8000/v1/completions"
-        self.headers = {"Content-Type": "application/json"}
-
-    def __call__(self, data):
-        prompt = data.get("inputs", "")
-        payload = {
-            "model": "TanyaVM/MistralAWQ2",
-            "prompt": prompt,
-            "max_tokens": 1500,
-        }
-
-        response = requests.post(self.api_url, json=payload, headers=self.headers)
-        result = response.json()
+import requests
+
+class EndpointHandler:
+    def __init__(self, path=""):
+        self.api_url = "http://vllm_generate/v1/completions"
+        self.headers = {"Content-Type": "application/json"}
+
+    def __call__(self, data):
+        prompt = data.get("inputs", "")
+        payload = {
+            "model": "TanyaVM/MistralAWQ2",
+            "prompt": prompt,
+            "max_tokens": 1500,
+        }
+
+        response = requests.post(self.api_url, json=payload, headers=self.headers)
+        result = response.json()
         return result