inference: load partial manifest, don't fully fail on unknown models

Author: ctheune

src/skvaider/inference/routers/manager.py

@@ -1,6 +1,6 @@
 import structlog
 import svcs
-from fastapi import APIRouter, HTTPException
+from fastapi import APIRouter
 from fastapi.responses import JSONResponse
 
 from skvaider.inference.manager import Manager
@@ -64,18 +64,18 @@ async def update_manifest(
 ) -> JSONResponse:
     """Update the manifest of models to load; the manager converges asynchronously."""
     manager = services.get(Manager)
-    models = {m.lower() for m in body.models}
-    unknown = models - manager.models.keys()
+    unknown = body.models - manager.models.keys()
+    models = body.models - unknown
+    manager.update_manifest(models, body.serial)
     if unknown:
-        log.error(
-            "manifest rejected: unknown models",
+        log.warning(
+            "manifest accepted with errors: unknown models",
             unknown=sorted(unknown),
-            requested=sorted(models),
-            configured=sorted(manager.models.keys()),
+            manifest=sorted(models),
+            known=sorted(manager.models.keys()),
         )
-        raise HTTPException(
-            status_code=500,
-            detail=f"Unknown models: {sorted(unknown)}",
+        return JSONResponse(
+            status_code=400,
+            content={"status": "unknown models", "unknown": sorted(unknown)},
         )
-    manager.update_manifest(models, body.serial)
     return JSONResponse(status_code=202, content={"status": "ok"})

src/skvaider/manifest.py

@@ -1,7 +1,7 @@
 import time
 from typing import Any, Literal
 
-from pydantic import GetCoreSchemaHandler
+from pydantic import GetCoreSchemaHandler, field_validator
 from pydantic_core import core_schema
 
 from skvaider.utils import RequestMethod, RequestModel, ResponseModel
@@ -121,3 +121,8 @@ class ManifestRequest(RequestModel[ManifestResponse]):
 
     models: set[str]
     serial: Serial
+
+    @field_validator("models", mode="after")
+    @classmethod
+    def normalize_models(cls, models: set[str]) -> set[str]:
+        return {m.lower() for m in models}