Log if cuda graph is used & extend cuda graph capture to cuda-graph-max-bs (#6201)

Co-authored-by: SangBin Cho <rkooo567@gmail.com>

test/srt/run_suite.py

@@ -101,8 +101,8 @@ suites = {
         # TestFile("test_deepep_intranode.py", 50),
         # TestFile("test_deepep_low_latency.py", 50),
         # TestFile("test_moe_deepep_eval_accuracy_large.py", 250),
+        # TestFile("test_disaggregation.py", 90),
         TestFile("test_local_attn.py", 250),
-        TestFile("test_disaggregation.py", 90),
         TestFile("test_full_deepseek_v3.py", 250),
         TestFile("test_pp_single_node.py", 150),
     ],

test/srt/test_eagle_infer.py

@@ -97,7 +97,9 @@ class TestEAGLEEngine(CustomTestCase):
 
         print(f"{engine.get_server_info()=}")
 
-        avg_spec_accept_length = engine.get_server_info()["avg_spec_accept_length"]
+        avg_spec_accept_length = engine.get_server_info()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 1.9)
 
@@ -296,7 +298,9 @@ class TestEAGLEServer(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.20)
 
         server_info = requests.get(self.base_url + "/get_server_info").json()
-        avg_spec_accept_length = server_info["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
 
         speculative_eagle_topk = server_info["speculative_eagle_topk"]

test/srt/test_fa3.py

@@ -111,7 +111,9 @@ class BaseFlashAttentionTest(CustomTestCase):
 
         if self.speculative_decode:
             server_info = requests.get(self.base_url + "/get_server_info")
-            avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+            avg_spec_accept_length = server_info.json()["internal_states"][0][
+                "avg_spec_accept_length"
+            ]
             print(f"{avg_spec_accept_length=}")
             self.assertGreater(avg_spec_accept_length, self.spec_decode_threshold)
 

test/srt/test_full_deepseek_v3.py

@@ -118,7 +118,9 @@ class TestDeepseekV3MTP(CustomTestCase):
         print(f"{metrics=}")
 
         server_info = requests.get(self.base_url + "/get_server_info")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
 
         if is_in_ci():

test/srt/test_mla_deepseek_v3.py

@@ -100,7 +100,9 @@ class TestDeepseekV3MTP(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.60)
 
         server_info = requests.get(self.base_url + "/get_server_info")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 2.5)
 
@@ -159,7 +161,9 @@ class TestDeepseekV3MTPWithDraft(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.60)
 
         server_info = requests.get(self.base_url + "/get_server_info")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 2.5)
 

test/srt/test_mla_flashinfer.py

@@ -158,7 +158,9 @@ class TestFlashinferMLAMTP(CustomTestCase):
 
         server_info = requests.get(self.base_url + "/get_server_info")
         print(f"{server_info=}")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 2.5)
 

test/srt/test_mla_int8_deepseek_v3.py

@@ -105,7 +105,9 @@ class TestDeepseekV3MTPChannelInt8(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.60)
 
         server_info = requests.get(self.base_url + "/get_server_info")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 2.5)
 
@@ -199,7 +201,9 @@ class TestDeepseekV3MTPBlockInt8(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.60)
 
         server_info = requests.get(self.base_url + "/get_server_info")
-        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
         print(f"{avg_spec_accept_length=}")
         self.assertGreater(avg_spec_accept_length, 2.5)
 

test/srt/test_srt_endpoint.py

@@ -492,9 +492,6 @@ class TestSRTEndpoint(CustomTestCase):
         max_total_num_tokens = response_json["max_total_num_tokens"]
         self.assertIsInstance(max_total_num_tokens, int)
 
-        attention_backend = response_json["attention_backend"]
-        self.assertIsInstance(attention_backend, str)
-
         version = response_json["version"]
         self.assertIsInstance(version, str)