Merge remote-tracking branch 'upstream/main' into HEAD

endpoints/OAI/router.py

@@ -22,6 +22,7 @@ from endpoints.OAI.utils.chat_completion import (
 )
 from endpoints.OAI.utils.completion import (
     generate_completion,
+    load_inline_model,
     stream_generate_completion,
 )
 from endpoints.OAI.utils.embeddings import get_embeddings
@@ -42,7 +43,7 @@ def setup():
 # Completions endpoint
 @router.post(
     "/v1/completions",
-    dependencies=[Depends(check_api_key), Depends(check_model_container)],
+    dependencies=[Depends(check_api_key)],
 )
 async def completion_request(
     request: Request, data: CompletionRequest
@@ -53,6 +54,18 @@ async def completion_request(
     If stream = true, this returns an SSE stream.
     """
 
+    if data.model:
+        inline_load_task = asyncio.create_task(load_inline_model(data.model, request))
+
+        await run_with_request_disconnect(
+            request,
+            inline_load_task,
+            disconnect_message=f"Model switch for generation {request.state.id} "
+            + "cancelled by user.",
+        )
+    else:
+        await check_model_container()
+
     model_path = model.container.model_dir
 
     if isinstance(data.prompt, list):
@@ -85,7 +98,7 @@ async def completion_request(
 # Chat completions endpoint
 @router.post(
     "/v1/chat/completions",
-    dependencies=[Depends(check_api_key), Depends(check_model_container)],
+    dependencies=[Depends(check_api_key)],
 )
 async def chat_completion_request(
     request: Request, data: ChatCompletionRequest
@@ -96,6 +109,11 @@ async def chat_completion_request(
     If stream = true, this returns an SSE stream.
     """
 
+    if data.model:
+        await load_inline_model(data.model, request)
+    else:
+        await check_model_container()
+
     if model.container.prompt_template is None:
         error_message = handle_request_error(
             "Chat completions are disabled because a prompt template is not set.",

endpoints/OAI/types/chat_completion.py

@@ -56,6 +56,7 @@ class ChatCompletionRequest(CommonCompletionRequest):
     add_generation_prompt: Optional[bool] = True
     template_vars: Optional[dict] = {}
     response_prefix: Optional[str] = None
+    model: Optional[str] = None
 
     # tools is follows the format OAI schema, functions is more flexible
     # both are available in the chat template.

endpoints/OAI/utils/completion.py

@@ -1,4 +1,8 @@
-"""Completion utilities for OAI server."""
+"""
+Completion utilities for OAI server.
+
+Also serves as a common module for completions and chat completions.
+"""
 
 import asyncio
 import pathlib
@@ -10,12 +14,14 @@ from typing import List, Union
 from loguru import logger
 
 from common import model
+from common.auth import get_key_permission
 from common.networking import (
     get_generator_error,
     handle_request_disconnect,
     handle_request_error,
     request_disconnect_loop,
 )
+from common.tabby_config import config
 from common.utils import unwrap
 from endpoints.OAI.types.completion import (
     CompletionRequest,
@@ -103,6 +109,50 @@ async def _stream_collector(
         await gen_queue.put(e)
 
 
+async def load_inline_model(model_name: str, request: Request):
+    """Load a model from the data.model parameter"""
+
+    # Return if the model container already exists and the model is fully loaded
+    if (
+        model.container
+        and model.container.model_dir.name == model_name
+        and model.container.model_loaded
+    ):
+        return
+
+    # Inline model loading isn't enabled or the user isn't an admin
+    if not get_key_permission(request) == "admin":
+        error_message = handle_request_error(
+            f"Unable to switch model to {model_name} because "
+            + "an admin key isn't provided",
+            exc_info=False,
+        ).error.message
+
+        raise HTTPException(401, error_message)
+
+    if not unwrap(config.model.get("inline_model_loading"), False):
+        logger.warning(
+            f"Unable to switch model to {model_name} because "
+            '"inline_model_loading" is not True in config.yml.'
+        )
+
+        return
+
+    model_path = pathlib.Path(unwrap(config.model.get("model_dir"), "models"))
+    model_path = model_path / model_name
+
+    # Model path doesn't exist
+    if not model_path.exists():
+        logger.warning(
+            f"Could not find model path {str(model_path)}. Skipping inline model load."
+        )
+
+        return
+
+    # Load the model
+    await model.load_model(model_path)
+
+
 async def stream_generate_completion(
     data: CompletionRequest, request: Request, model_path: pathlib.Path
 ):