added support for additional more user-friendly interfaces, improved some part of the application loading process to make it a bit simpler

samples/models/dummy/config.json

@@ -2,10 +2,14 @@
   "type": "python",
   "tags": ["dummy"],
   "file": "model.py",
+  "interface": "chat",
 
-  "output_type": "video/mp4",
+  "summary": "Echo model",
+  "description": "The most basic example model, simply echo the input",
 
   "inputs": {
-    "file": {"type": "file"}
-  }
+    "messages": {"type": "list[dict]", "default": "[{\"role\": \"user\", \"content\": \"who are you ?\"}]"}
+  },
+
+  "output_type": "text/markdown"
 }

samples/models/dummy/model.py

@@ -7,5 +7,5 @@ def load(model) -> None:
 def unload(model) -> None:
     pass
 
-async def infer(model, file) -> typing.AsyncIterator[bytes]:
-    yield await file.read()
+async def infer(model, messages: list[dict]) -> typing.AsyncIterator[bytes]:
+    yield messages[-1]["content"].encode("utf-8")

samples/models/python-bert-1/model.py

@@ -16,7 +16,7 @@ def unload(model) -> None:
     model.model = None
     model.tokenizer = None
 
-def infer(model, prompt: str) -> typing.Iterator[bytes]:
+async def infer(model, prompt: str) -> typing.AsyncIterator[bytes]:
     inputs = model.tokenizer(prompt, return_tensors="pt")
 
     with torch.no_grad():

samples/models/python-bert-2/model.py

@@ -16,7 +16,7 @@ def unload(model) -> None:
     model.model = None
     model.tokenizer = None
 
-def infer(model, prompt: str) -> typing.Iterator[bytes]:
+async def infer(model, prompt: str) -> typing.AsyncIterator[bytes]:
     inputs = model.tokenizer(prompt, return_tensors="pt")
 
     with torch.no_grad():