Add return hidden state in the native API (#3897)

Co-authored-by: Beichen-Ma <mabeichen12@gmail.com>
Co-authored-by: Chayenne <zhaochen20@outlook.com>

examples/runtime/engine/hidden_states.py

@@ -2,7 +2,9 @@
 Usage:
 python hidden_states.py
 
-Note that we are actively working on moving return_hidden_states to the sampling_params.
+Note that each time you change the `return_hidden_states` parameter,
+the cuda graph will be recaptured, which might lead to a performance hit.
+So avoid getting hidden states and completions alternately.
 """
 
 import sglang as sgl
@@ -18,10 +20,14 @@ def main():
     # Create an LLM.
     llm = sgl.Engine(
         model_path="Alibaba-NLP/gte-Qwen2-1.5B-instruct",
-        return_hidden_states=True,
     )
 
-    sampling_params = {"temperature": 0.8, "top_p": 0.95, "max_new_tokens": 10}
+    sampling_params = {
+        "temperature": 0.8,
+        "top_p": 0.95,
+        "max_new_tokens": 10,
+        "return_hidden_states": True,
+    }
 
     outputs = llm.generate(prompts, sampling_params=sampling_params)
     for prompt, output in zip(prompts, outputs):