xinference 1.5.0__py3-none-any.whl → 1.5.0.post2__py3-none-any.whl

xinference/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2025-04-19T20:32:22+0800",
+ "date": "2025-04-21T17:53:44+0800",
  "dirty": false,
  "error": null,
- "full-revisionid": "ee8d025e1c046b22b3b148e5e97c0e107c979ee3",
- "version": "1.5.0"
+ "full-revisionid": "a5d4be9f970137bde1d402420f71961826392224",
+ "version": "1.5.0.post2"
 }
 '''  # END VERSION_JSON
 

xinference/core/supervisor.py

@@ -1356,7 +1356,12 @@ class SupervisorActor(xo.StatelessActor):
         return model_uid
 
     async def get_launch_builtin_model_progress(self, model_uid: str) -> float:
-        info = self._model_uid_to_replica_info[model_uid]
+        try:
+            info = self._model_uid_to_replica_info[model_uid]
+        except KeyError:
+            # Not launched perhaps, just return 0.0 to prevent error
+            return 0.0
+
         all_progress = 0.0
         i = 0
         for rep_model_uid in iter_replica_model_uid(model_uid, info.replica):
@@ -1370,7 +1375,11 @@ class SupervisorActor(xo.StatelessActor):
         return all_progress / i if i > 0 else 0.0
 
     async def cancel_launch_builtin_model(self, model_uid: str):
-        info = self._model_uid_to_replica_info[model_uid]
+        try:
+            info = self._model_uid_to_replica_info[model_uid]
+        except KeyError:
+            raise RuntimeError(f"Model {model_uid} has not been launched yet")
+
         coros = []
         for i, rep_model_uid in enumerate(
             iter_replica_model_uid(model_uid, info.replica)

xinference/model/llm/llm_family.json

@@ -7946,7 +7946,9 @@
     "virtualenv": {
       "packages": [
         "git+https://github.com/huggingface/transformers@v4.51.3-Qwen2.5-Omni-preview",
-        "numpy==1.26.4"
+        "numpy==1.26.4",
+        "qwen_omni_utils",
+        "soundfile"
       ]
     }
   },

xinference/model/llm/llm_family_modelscope.json

@@ -5704,7 +5704,9 @@
     "virtualenv": {
       "packages": [
         "git+https://github.com/huggingface/transformers@v4.51.3-Qwen2.5-Omni-preview",
-        "numpy==1.26.4"
+        "numpy==1.26.4",
+        "qwen_omni_utils",
+        "soundfile"
       ]
     }
   },

xinference/model/llm/transformers/gemma3.py

@@ -128,7 +128,12 @@ class Gemma3ChatModel(PytorchChatModel):
         ).to(self._device)
         input_len = inputs["input_ids"].shape[-1]
 
-        generation = self._model.generate(**inputs, do_sample=False)
+        generation = self._model.generate(
+            **inputs,
+            do_sample=False,
+            max_new_tokens=config.get("max_tokens", 512),
+            temperature=config.get("temperature", 1),
+        )
         generation = generation[0][input_len:]
 
         decoded = self._processor.decode(generation, skip_special_tokens=True)
@@ -159,7 +164,11 @@ class Gemma3ChatModel(PytorchChatModel):
 
         def model_generate():
             try:
-                return self._model.generate(**gen_kwargs)
+                return self._model.generate(
+                    **gen_kwargs,
+                    max_new_tokens=config.get("max_tokens", 512),
+                    temperature=config.get("temperature", 1),
+                )
             except Exception:
                 nonlocal error
                 error = sys.exc_info()

xinference/web/ui/build/asset-manifest.json

@@ -1,14 +1,14 @@
 {
   "files": {
     "main.css": "./static/css/main.0f6523be.css",
-    "main.js": "./static/js/main.58bd483c.js",
+    "main.js": "./static/js/main.4b67a723.js",
     "static/media/icon.webp": "./static/media/icon.4603d52c63041e5dfbfd.webp",
     "index.html": "./index.html",
     "main.0f6523be.css.map": "./static/css/main.0f6523be.css.map",
-    "main.58bd483c.js.map": "./static/js/main.58bd483c.js.map"
+    "main.4b67a723.js.map": "./static/js/main.4b67a723.js.map"
   },
   "entrypoints": [
     "static/css/main.0f6523be.css",
-    "static/js/main.58bd483c.js"
+    "static/js/main.4b67a723.js"
   ]
 }

xinference/web/ui/build/index.html

@@ -1 +1 @@
-<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.58bd483c.js"></script><link href="./static/css/main.0f6523be.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>
+<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.4b67a723.js"></script><link href="./static/css/main.0f6523be.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>