[Fix] Fix and rename flashmla CI test (#6045)

test/srt/run_suite.py

@@ -51,7 +51,7 @@ suites = {
         TestFile("test_mla_int8_deepseek_v3.py", 389),
         TestFile("test_mla_flashinfer.py", 395),
         TestFile("test_mla_fp8.py", 153),
-        TestFile("test_flash_mla_attention_backend.py", 300),
+        TestFile("test_flashmla.py", 300),
         TestFile("test_no_chunked_prefill.py", 108),
         TestFile("test_no_overlap_scheduler.py", 216),
         TestFile("test_openai_server.py", 149),

test/srt/test_flash_mla_attention_backend.py

@@ -1,64 +0,0 @@
-"""
-Usage:
-python3 -m unittest test_flash_mla_attention_backend.TestFlashMLAAttnBackend.test_mmlu
-"""
-
-import unittest
-from types import SimpleNamespace
-
-from sglang.srt.utils import kill_process_tree
-from sglang.test.run_eval import run_eval
-from sglang.test.test_utils import (
-    DEFAULT_MLA_MODEL_NAME_FOR_TEST,
-    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-    DEFAULT_URL_FOR_TEST,
-    is_in_ci,
-    popen_launch_server,
-    run_bench_one_batch,
-)
-
-
-class TestFlashMLAAttnBackend(unittest.TestCase):
-    def test_latency(self):
-        output_throughput = run_bench_one_batch(
-            DEFAULT_MLA_MODEL_NAME_FOR_TEST,
-            [
-                "--attention-backend",
-                "flashmla",
-                "--enable-torch-compile",
-                "--cuda-graph-max-bs",
-                "16",
-                "--trust-remote-code",
-            ],
-        )
-
-        if is_in_ci():
-            self.assertGreater(output_throughput, 153)
-
-    def test_mmlu(self):
-        model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
-        base_url = DEFAULT_URL_FOR_TEST
-        process = popen_launch_server(
-            model,
-            base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--attention-backend", "flashmla", "--trust-remote-code"],
-        )
-
-        try:
-            args = SimpleNamespace(
-                base_url=base_url,
-                model=model,
-                eval_name="mmlu",
-                num_examples=64,
-                num_threads=32,
-            )
-
-            metrics = run_eval(args)
-            self.assertGreaterEqual(metrics["score"], 0.2)
-        finally:
-            kill_process_tree(process.pid)
-
-
-if __name__ == "__main__":
-    unittest.main()

test/srt/test_flashmla.py

@@ -0,0 +1,86 @@
+"""
+Usage:
+python3 test/srt/test_flashmla.py
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import torch
+
+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.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST_MLA,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    run_bench_one_batch,
+)
+
+
+class TestFlashMLAAttnBackend(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = ["--trust-remote-code"]
+        if torch.cuda.is_available() and torch.version.cuda:
+            other_args.extend(
+                [
+                    "--enable-torch-compile",
+                    "--cuda-graph-max-bs",
+                    "2",
+                    "--attention-backend",
+                    "flashmla",
+                ]
+            )
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(metrics)
+
+        self.assertGreater(metrics["accuracy"], 0.62)
+
+
+class TestFlashMLAAttnLatency(unittest.TestCase):
+    def test_latency(self):
+        output_throughput = run_bench_one_batch(
+            DEFAULT_MODEL_NAME_FOR_TEST_MLA,
+            [
+                "--attention-backend",
+                "flashmla",
+                "--enable-torch-compile",
+                "--cuda-graph-max-bs",
+                "16",
+                "--trust-remote-code",
+            ],
+        )
+
+        if is_in_ci():
+            self.assertGreater(output_throughput, 100)
+
+
+if __name__ == "__main__":
+    unittest.main()