Let reward model take text inputs instead of message lists (#1907)

Co-authored-by: Kyle Corbitt <kyle@corbt.com>

python/sglang/test/runners.py

@@ -273,6 +273,7 @@ class SRTRunner:
             disable_cuda_graph=disable_cuda_graph,
             disable_radix_cache=disable_radix_cache,
         )
+        self.tokenizer = get_tokenizer(model_path)
 
     def forward(
         self,
@@ -366,7 +367,7 @@ class SRTRunner:
                 return ModelOutput(embed_logits=logits)
             else:
                 scores = [x["embedding"][0] for x in response]
-                return ModelOutput(scores=logits)
+                return ModelOutput(scores=scores)
 
     def __enter__(self):
         return self