fix:

main.py

@@ -9,6 +9,7 @@ from api.routers import health
 from infra.logger_structlog import StructLogger
 
 from contextlib import asynccontextmanager
+import os
 import mlflow
 import torch
 import redis.asyncio as aioredis
@@ -17,13 +18,17 @@ import dagshub
 from fastapi import FastAPI
 
 
+settings = AppConfig()
+if settings.dagshub_user_token:
+    os.environ["DAGSHUB_USER_TOKEN"] = settings.dagshub_user_token
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """
     This is on_event("startup") new alternative, Make sure you load models here.
     """
 
-    settings = AppConfig()
     logger = StructLogger(settings=settings)
     # Using this way to can store data. it is acts as a dict which holds instances
     app.state.logger = logger
@@ -34,13 +39,16 @@ async def lifespan(app: FastAPI):
     # asyncio.create_task(log_system_metrics(logger, logger_interval_sec=settings.intervals.system_metrics_seconds))
 
     detection_model_path = hf_fetch_model(
-        repo_id="Ultralytics/YOLO26", filename=settings.yolo.model_name
+        repo_id="Ultralytics/YOLO26",
+        filename=settings.yolo.model_name,
+        cache_dir=settings.hf_cache_dir,
     )
     app.state.detection_model = YOLO_Detector(detection_model_path)
 
     depth_model_path = hf_fetch_model(
         repo_id="depth-anything/Depth-Anything-V2-Small",
         filename=settings.depth.model_name,
+        cache_dir=settings.hf_cache_dir,
     )
     app.state.depth_model = DepthAnything(
         encoder=settings.depth.encoder,
@@ -49,7 +57,9 @@ async def lifespan(app: FastAPI):
     )
 
     safety_detection_path = hf_fetch_model(
-        repo_id="e1250/safety_detection", filename=settings.security_detector.model_name
+        repo_id="e1250/safety_detection",
+        filename=settings.security_detector.model_name,
+        cache_dir=settings.hf_cache_dir,
     )
     app.state.safety_detection_model = YOLO_Detector(safety_detection_path)