Tiny apply gsm8k mixin to ngram test (#15606)

python/sglang/test/kits/gsm8k_accuracy_kit.py

@@ -1,13 +1,18 @@
 from types import SimpleNamespace
+from typing import Optional
+
+import requests
 
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_gsm8k
-from sglang.test.test_utils import CustomTestCase
 
 
 class GSM8KMixin:
-    accuracy: float
+    gsm8k_accuracy_thres: float
+    gsm8k_accept_length_thres: Optional[float] = None
+
+    def test_gsm8k(self):
+        requests.get(self.base_url + "/flush_cache")
 
-    def test_gsm8k(self: CustomTestCase):
         args = SimpleNamespace(
             num_shots=5,
             data_path=None,
@@ -19,4 +24,12 @@ class GSM8KMixin:
         )
         metrics = run_eval_gsm8k(args)
         print(f"{metrics=}")
-        self.assertGreaterEqual(metrics["accuracy"], self.accuracy)
+        self.assertGreaterEqual(metrics["accuracy"], self.gsm8k_accuracy_thres)
+
+        if self.gsm8k_accept_length_thres is not None:
+            server_info = requests.get(self.base_url + "/server_info")
+            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.gsm8k_accept_length_thres)

test/registered/spec/test_ngram_speculative_decoding.py

@@ -1,11 +1,9 @@
 import unittest
-from types import SimpleNamespace
-
-import requests
 
 from sglang.srt.environ import envs
 from sglang.srt.utils import kill_process_tree
-from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
 from sglang.test.test_utils import (
     DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -14,6 +12,8 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+register_cuda_ci(est_time=117, suite="stage-b-test-small-1-gpu")
+
 GSM_DATASET_PATH = None
 
 
@@ -31,12 +31,11 @@ DEFAULT_SERVER_ARGS = [
 ]
 
 
-class TestNgramSpeculativeDecodingBase(CustomTestCase):
-
+class TestNgramSpeculativeDecodingBase(GSM8KMixin, CustomTestCase):
     model = DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST
     base_url = DEFAULT_URL_FOR_TEST
-    accuracy_threshold = 0.79  # derived tests need to override this
-    spec_decode_threshold = 1.8  # derived spec decoding tests need to override this
+    gsm8k_accuracy_thres = 0.79  # derived tests need to override this
+    gsm8k_accept_length_thres = 1.8  # derived spec decoding tests need to override this
 
     @classmethod
     def get_server_args(cls):
@@ -61,32 +60,6 @@ class TestNgramSpeculativeDecodingBase(CustomTestCase):
     def tearDownClass(cls):
         kill_process_tree(cls.process.pid)
 
-    def test_gsm8k(self):
-        requests.get(self.base_url + "/flush_cache")
-
-        args = SimpleNamespace(
-            num_shots=4,
-            num_questions=100,
-            max_new_tokens=512,
-            parallel=128,
-            host="http://127.0.0.1",
-            port=int(self.base_url.split(":")[-1]),
-            data_path=GSM_DATASET_PATH,
-        )
-        metrics = run_eval_few_shot_gsm8k(args)
-        print(f"{metrics=}")
-
-        # Use the appropriate metric key based on the test class
-        metric_key = "accuracy"
-        self.assertGreater(metrics[metric_key], self.accuracy_threshold)
-
-        server_info = requests.get(self.base_url + "/get_server_info")
-        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)
-
 
 class TestNgramSpeculativeDecodingTriton(TestNgramSpeculativeDecodingBase):
 

test/srt/models/test_ministral3_models.py

@@ -9,7 +9,7 @@ MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
 
 
 class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
-    accuracy = 0.6
+    gsm8k_accuracy_thres = 0.6
     model = MODEL
     other_args = ["--trust-remote-code"]
 

test/srt/models/test_nvidia_nemotron_nano_v2.py

@@ -7,19 +7,19 @@ from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
 class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
-    accuracy = 0.87
+    gsm8k_accuracy_thres = 0.87
     other_args = ["--max-mamba-cache-size", "256"]
 
 
 class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
-    accuracy = 0.87
+    gsm8k_accuracy_thres = 0.87
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8"
     other_args = ["--max-mamba-cache-size", "256"]
 
 
 @unittest.skipIf(not is_blackwell(), "NVFP4 only supported on blackwell")
 class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
-    accuracy = 0.855
+    gsm8k_accuracy_thres = 0.855
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-NVFP4"
     other_args = ["--max-mamba-cache-size", "256"]
 
@@ -29,7 +29,7 @@ class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
     "with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
 )
 class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase):
-    accuracy = 0.87
+    gsm8k_accuracy_thres = 0.87
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
     other_args = [
         "--speculative-algorithm",
@@ -60,7 +60,7 @@ class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase)
 class TestNvidiaNemotronNanoV2SpeculativeDecodingBF16Cache(
     GSM8KMixin, DefaultServerBase
 ):
-    accuracy = 0.87
+    gsm8k_accuracy_thres = 0.87
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
     other_args = [
         "--speculative-algorithm",

test/srt/models/test_nvidia_nemotron_nano_v2_vl.py

@@ -9,7 +9,7 @@ MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
 
 
 class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, DefaultServerBase):
-    accuracy = 0.87
+    gsm8k_accuracy_thres = 0.87
     model = MODEL
     other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]
 

test/srt/run_suite.py

@@ -79,7 +79,6 @@ suites = {
         TestFile("test_model_hooks.py", 6),
         TestFile("test_modelopt_loader.py", 11),
         TestFile("test_multi_tokenizer.py", 230),
-        TestFile("test_ngram_speculative_decoding.py", 177),
         TestFile("test_no_chunked_prefill.py", 108),
         TestFile("test_no_overlap_scheduler.py", 217),
         TestFile("test_original_logprobs.py", 41),