Arena as fisrt tab

app.py

@@ -403,21 +403,21 @@ class GradioApp:
         return chat, model_text
 
     def _build_arena_with_launched(self, models: List[Tuple[str, ModelSpec]]):
-        with gr.Box():
-            with gr.Row():
-                chat_and_text = [
-                    self._build_single_with_launched(models, i)
-                    for i in range(self._gladiator_num)
-                ]
-                chats = [c[0] for c in chat_and_text]
-                texts = [c[1] for c in chat_and_text]
+        chat_and_text = []
+        with gr.Row():
+            for i in range(self._gladiator_num):
+                with gr.Column():
+                    chat_and_text.append(self._build_single_with_launched(models, i))
 
-            msg = gr.Textbox(label=self._locale("Input"))
+        chats = [c[0] for c in chat_and_text]
+        texts = [c[1] for c in chat_and_text]
 
-            def update_message(text_in: str):
-                return "", text_in, text_in
+        msg = gr.Textbox(label=self._locale("Input"))
 
-            msg.submit(update_message, inputs=[msg], outputs=[msg] + texts)
+        def update_message(text_in: str):
+            return "", text_in, text_in
+
+        msg.submit(update_message, inputs=[msg], outputs=[msg] + texts)
 
         gr.ClearButton(components=[msg] + chats + texts)
 
@@ -425,6 +425,9 @@ class GradioApp:
         if self._use_launched_model:
             models = self._api.list_models()
             with gr.Blocks() as blocks:
+                if len(models) >= 2:
+                    with gr.Tab(self._locale("Arena")):
+                        self._build_arena_with_launched(models)
                 with gr.Tab(self._locale("Chat")):
                     chat, model_text = self._build_single_with_launched(models, 0)
                     msg = gr.Textbox(label=self._locale("Input"))
@@ -434,9 +437,6 @@ class GradioApp:
 
                     msg.submit(update_message, inputs=[msg], outputs=[msg, model_text])
                     gr.ClearButton(components=[chat, msg, model_text])
-                if len(models) > 2:
-                    with gr.Tab(self._locale("Arena")):
-                        self._build_arena_with_launched(models)
         else:
             with gr.Blocks() as blocks:
                 with gr.Tab(self._locale("Chat")):
@@ -450,16 +450,23 @@ class GradioApp:
 async def launch_xinference():
     import xoscar as xo
     from xinference.core.service import SupervisorActor
+    from xinference.core.api import AsyncSupervisorAPI
     from xinference.deploy.worker import start_worker_components
 
     pool = await xo.create_actor_pool(address="0.0.0.0", n_process=0)
+    supervisor_address = pool.external_address
     await xo.create_actor(
-        SupervisorActor, address=pool.external_address, uid=SupervisorActor.uid()
+        SupervisorActor, address=supervisor_address, uid=SupervisorActor.uid()
     )
     await start_worker_components(
-        address=pool.external_address, supervisor_address=pool.external_address
+        address=supervisor_address, supervisor_address=supervisor_address
     )
-    gradio_block = GradioApp(pool.external_address).build()
+    api = AsyncSupervisorAPI(supervisor_address)
+    supported_models = ["orca", "chatglm2", "chatglm", "vicuna-v1.3"]
+    for model in supported_models:
+        await api.launch_model(str(uuid.uuid4()), model)
+
+    gradio_block = GradioApp(supervisor_address, use_launched_model=True).build()
     gradio_block.launch()
 
 

requirements.txt

@@ -1,4 +1,4 @@
-xinference
+xinference @ git+https://github.com/aresnow1/inference.git@bugfix/model-lock
 xoscar
 chatglm-cpp
 llama-cpp-python