[CI] balance unit tests (#1977)

python/sglang/srt/server.py

@@ -114,9 +114,16 @@ async def health() -> Response:
 @app.get("/health_generate")
 async def health_generate(request: Request) -> Response:
     """Check the health of the inference server by generating one token."""
-    gri = GenerateReqInput(
-        text="s", sampling_params={"max_new_tokens": 1, "temperature": 0.7}
-    )
+
+    if tokenizer_manager.is_generation:
+        gri = GenerateReqInput(
+            input_ids=[0], sampling_params={"max_new_tokens": 1, "temperature": 0.7}
+        )
+    else:
+        gri = EmbeddingReqInput(
+            input_ids=[0], sampling_params={"max_new_tokens": 1, "temperature": 0.7}
+        )
+
     try:
         async for _ in tokenizer_manager.generate_request(gri, request):
             break

python/sglang/test/test_utils.py

@@ -442,7 +442,7 @@ def popen_launch_server(
                 "Content-Type": "application/json; charset=utf-8",
                 "Authorization": f"Bearer {api_key}",
             }
-            response = requests.get(f"{base_url}/v1/models", headers=headers)
+            response = requests.get(f"{base_url}/health_generate", headers=headers)
             if response.status_code == 200:
                 return process
         except requests.RequestException:
@@ -637,8 +637,8 @@ def calculate_rouge_l(output_strs_list1, output_strs_list2):
     return rouge_l_scores
 
 
-STDOUT_FILENAME = "stdout.txt"
 STDERR_FILENAME = "stderr.txt"
+STDOUT_FILENAME = "stdout.txt"
 
 
 def read_output(output_lines):