test_mla_deepseek_v3.py

import os
import unittest
from types import SimpleNamespace

import requests

from sglang.srt.utils import is_cuda, is_hip, 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,
    CustomTestCase,
    is_in_ci,
    popen_launch_server,
)


class TestMLADeepseekV3(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
        if is_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)


@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestMLADeepseekV3DisableFusedFunc(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        os.environ["SGLANG_CI_DISABLE_MOE_FUSED_FUNC"] = "1"
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
        if is_cuda():
            other_args.extend(["--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)


@unittest.skipIf(is_hip(), "FA is not available.")
class TestMLADeepseekV3Fa3Fp8Kvcache(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--chunked-prefill-size",
            "256",
            "--kv-cache-dtype",
            "fp8_e4m3",
        ]
        if is_cuda():
            other_args.extend(["--attention-backend", "fa3"])
        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 TestDeepseekV3MTP(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--cuda-graph-max-bs",
            "2",
            "--disable-radix",
            "--enable-torch-compile",
            "--torch-compile-max-bs",
            "1",
            "--speculative-algorithm",
            "EAGLE",
            "--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()["internal_states"][0][
            "avg_spec_accept_length"
        ]
        print(f"{avg_spec_accept_length=}")
        self.assertGreater(avg_spec_accept_length, 2.5)


if __name__ == "__main__":
    unittest.main()