diff --git a/docs/user_guide/install.md b/docs/user_guide/install.md
index 87854649..02a30343 100644
--- a/docs/user_guide/install.md
+++ b/docs/user_guide/install.md
@@ -82,12 +82,10 @@ For Linux+GPU devices, Parallax provides a docker environment for quick setup. C
 
 Run a docker container as below. Please note that generally the argument ```--gpus all``` is necessary for the docker to run on GPUs.
 ```sh
-# For Blackwell
-docker run -it --gpus all --network host gradientservice/parallax:latest-blackwell bash
-# For Ampere/Hopper
-docker run -it --gpus all --network host gradientservice/parallax:latest-hopper bash
+# For Blackwell/Ampere/Hopper
+docker run -it --gpus all --network host gradientservice/parallax:latest bash
 # For DGX Spark
-docker run -it --gpus all --network host gradientservice/parallax:spark-spark bash
+docker run -it --gpus all --network host gradientservice/parallax:latest-spark bash
 ```
 The container starts under parallax workspace and you should be able to run parallax directly.
 
diff --git a/src/parallax/launch.py b/src/parallax/launch.py
index f26f2e60..56589445 100644
--- a/src/parallax/launch.py
+++ b/src/parallax/launch.py
@@ -136,7 +136,9 @@
             )
             args.start_layer = gradient_server.block_start_index
             args.end_layer = gradient_server.block_end_index
-            args.model_path = gradient_server.model_name
+            # Only read model_name from scheduler if model_path is not set, so we can use local path as model_path
+            if args.model_path is None:
+                args.model_path = gradient_server.model_name
             args.tp_size = gradient_server.tp_size
 
             logger.debug(
diff --git a/src/scheduling/request_routing.py b/src/scheduling/request_routing.py
index e4d5e09b..b478607f 100644
--- a/src/scheduling/request_routing.py
+++ b/src/scheduling/request_routing.py
@@ -430,7 +430,7 @@ def find_optimal_path(self, nodes: List[Node], num_layers: int) -> Tuple[List[st
                 prev = node
             self._rr_cursor += 1
             attempts += 1
-            if viable:
+            if viable and total_latency != float("inf"):
                 return candidate_ids, total_latency
             # Attempt a one-shot repair if the selected pipeline is not viable
             repaired = self._attempt_repair_pipeline(candidate_ids, nodes, num_layers)