add dsv3 int8 test (#4705)

test/srt/run_suite.py

@@ -41,6 +41,7 @@ suites = {
         TestFile("test_metrics.py", 32),
         TestFile("test_mla.py", 92),
         TestFile("test_mla_deepseek_v3.py", 221),
+        TestFile("test_mla_int8_deepseek_v3.py", 421),
         TestFile("test_mla_flashinfer.py", 395),
         TestFile("test_mla_fp8.py", 93),
         TestFile("test_no_chunked_prefill.py", 126),

test/srt/test_mla_int8_deepseek_v3.py

+import unittest
+from types import SimpleNamespace
+
+import requests
+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_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestMLADeepseekV3ChannelInt8(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "sgl-project/sglang-ci-dsv3-channel-int8-test"
+        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"])
+        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 TestDeepseekV3MTPChannelInt8(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "sgl-project/sglang-ci-dsv3-channel-int8-test"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = ["--trust-remote-code"]
+        if torch.cuda.is_available() and torch.version.cuda:
+            other_args.extend(
+                [
+                    "--cuda-graph-max-bs",
+                    "2",
+                    "--disable-radix",
+                    "--enable-torch-compile",
+                    "--torch-compile-max-bs",
+                    "1",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-draft",
+                    "sgl-project/sglang-ci-dsv3-channel-int8-test-NextN",
+                    "--speculative-num-steps",
+                    "2",
+                    "--speculative-eagle-topk",
+                    "4",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                ]
+            )
+        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):
+        requests.get(self.base_url + "/flush_cache")
+
+        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"], 0.60)
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        print(f"{avg_spec_accept_length=}")
+        self.assertGreater(avg_spec_accept_length, 2.5)
+
+
+class TestMLADeepseekV3BlockInt8(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "sgl-project/sglang-ci-dsv3-block-int8-test"
+        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"])
+        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 TestDeepseekV3MTPBlockInt8(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "sgl-project/sglang-ci-dsv3-block-int8-test"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = ["--trust-remote-code"]
+        if torch.cuda.is_available() and torch.version.cuda:
+            other_args.extend(
+                [
+                    "--cuda-graph-max-bs",
+                    "2",
+                    "--disable-radix",
+                    "--enable-torch-compile",
+                    "--torch-compile-max-bs",
+                    "1",
+                    "--speculative-algorithm",
+                    "EAGLE",
+                    "--speculative-draft",
+                    "sgl-project/sglang-ci-dsv3-block-int8-test-NextN",
+                    "--speculative-num-steps",
+                    "2",
+                    "--speculative-eagle-topk",
+                    "4",
+                    "--speculative-num-draft-tokens",
+                    "4",
+                ]
+            )
+        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):
+        requests.get(self.base_url + "/flush_cache")
+
+        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"], 0.60)
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["avg_spec_accept_length"]
+        print(f"{avg_spec_accept_length=}")
+        self.assertGreater(avg_spec_accept_length, 2.5)
+
+
+if __name__ == "__main__":
+    unittest.main()