Fix torch.compile for MoE (#2033)

test/srt/run_suite.py

@@ -27,6 +27,7 @@ suites = {
         "test_srt_engine.py",
         "test_srt_endpoint.py",
         "test_torch_compile.py",
+        "test_torch_compile_moe.py",
         "test_torchao.py",
         "test_triton_attention_kernels.py",
         "test_triton_attention_backend.py",

test/srt/test_data_parallelism.py

@@ -40,7 +40,7 @@ class TestDataParallelism(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.65
+        self.assertGreaterEqual(metrics["score"], 0.65)
 
     def test_update_weight(self):
         response = requests.post(

test/srt/test_double_sparsity.py

@@ -55,7 +55,7 @@ class TestDoubleSparsity(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.65
+        self.assertGreaterEqual(metrics["score"], 0.65)
 
 
 if __name__ == "__main__":

test/srt/test_eval_accuracy_mini.py

@@ -35,7 +35,7 @@ class TestEvalAccuracyMini(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.65
+        self.assertGreaterEqual(metrics["score"], 0.65)
 
 
 if __name__ == "__main__":

test/srt/test_retract_decode.py

@@ -34,7 +34,7 @@ class TestRetractDecode(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.65
+        self.assertGreaterEqual(metrics["score"], 0.65)
 
 
 if __name__ == "__main__":

test/srt/test_torch_compile.py

@@ -39,7 +39,7 @@ class TestTorchCompile(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.65
+        self.assertGreaterEqual(metrics["score"], 0.65)
 
     def run_decode(self, max_new_tokens):
         response = requests.post(
@@ -49,8 +49,8 @@ class TestTorchCompile(unittest.TestCase):
                 "sampling_params": {
                     "temperature": 0,
                     "max_new_tokens": max_new_tokens,
+                    "ignore_eos": True,
                 },
-                "ignore_eos": True,
             },
         )
         return response.json()
@@ -66,7 +66,7 @@ class TestTorchCompile(unittest.TestCase):
         print(res["text"])
         throughput = max_tokens / (tok - tic)
         print(f"Throughput: {throughput} tokens/s")
-        assert throughput >= 152
+        self.assertGreaterEqual(throughput, 152)
 
 
 if __name__ == "__main__":

test/srt/test_torch_compile_moe.py

@@ -0,0 +1,73 @@
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_child_process
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestTorchCompile(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=["--enable-torch-compile", "--torch-compile-max-bs", "1"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_child_process(cls.process.pid, include_self=True)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(metrics["score"], 0.50)
+
+    def run_decode(self, max_new_tokens):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": max_new_tokens,
+                    "ignore_eos": True,
+                },
+            },
+        )
+        return response.json()
+
+    def test_throughput(self):
+        import time
+
+        max_tokens = 256
+
+        tic = time.time()
+        res = self.run_decode(max_tokens)
+        tok = time.time()
+        print(f"{res=}")
+        throughput = max_tokens / (tok - tic)
+        print(f"Throughput: {throughput} tokens/s")
+        self.assertGreaterEqual(throughput, 290)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_triton_attention_backend.py

@@ -48,7 +48,7 @@ class TestTritonAttnBackend(unittest.TestCase):
             )
 
             metrics = run_eval(args)
-            assert metrics["score"] >= 0.65
+            self.assertGreaterEqual(metrics["score"], 0.65)
         finally:
             kill_child_process(process.pid, include_self=True)