/test/registered/disaggregation/test_disaggregation_pp.py

import time
import unittest
from types import SimpleNamespace

from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.server_fixtures.disaggregation_fixture import (
    PDDisaggregationServerBase,
)
from sglang.test.test_utils import (
    DEFAULT_MODEL_NAME_FOR_TEST,
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    popen_launch_pd_server,
    try_cached_model,
)

register_cuda_ci(est_time=216, stage="base-c", runner_config="8-gpu-h20")


class TestDisaggregationPrefillPPAccuracy(PDDisaggregationServerBase):
    @classmethod
    def setUpClass(cls):
        super().setUpClass()
        cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST)

        # Non blocking start servers
        cls.start_prefill()
        cls.start_decode()

        # Block until both
        cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
        cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)

        cls.launch_lb()

    @classmethod
    def start_prefill(cls):
        prefill_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "prefill",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--pp-size",
            "2",
            "--disable-overlap-schedule",
        ]
        prefill_args += cls.transfer_backend + cls.rdma_devices
        cls.process_prefill = popen_launch_pd_server(
            cls.model,
            cls.prefill_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=prefill_args,
        )

    @classmethod
    def start_decode(cls):
        decode_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "decode",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--base-gpu-id",
            "4",
        ]
        decode_args += cls.transfer_backend + cls.rdma_devices
        cls.process_decode = popen_launch_pd_server(
            cls.model,
            cls.decode_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=decode_args,
        )

    def test_gsm8k(self):
        args = SimpleNamespace(
            base_url=self.base_url,
            model=self.model,
            eval_name="gsm8k",
            api="completion",
            max_tokens=512,
            num_examples=200,
            num_threads=128,
        )
        metrics = run_eval(args)
        print(f"{metrics=}")

        self.assertGreater(metrics["score"], 0.24)
        # Wait a little bit so that the memory check happens.
        time.sleep(5)


class TestDisaggregationPrefillPPDynamicChunkAccuracy(PDDisaggregationServerBase):
    @classmethod
    def setUpClass(cls):
        super().setUpClass()
        cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST)

        # Non blocking start servers
        cls.start_prefill()
        cls.start_decode()

        # Block until both
        cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
        cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)

        cls.launch_lb()

    @classmethod
    def start_prefill(cls):
        prefill_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "prefill",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--pp-size",
            "2",
            "--disable-overlap-schedule",
            "--enable-dynamic-chunking",
        ]
        prefill_args += cls.transfer_backend + cls.rdma_devices
        cls.process_prefill = popen_launch_pd_server(
            cls.model,
            cls.prefill_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=prefill_args,
        )

    @classmethod
    def start_decode(cls):
        decode_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "decode",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--base-gpu-id",
            "4",
        ]
        decode_args += cls.transfer_backend + cls.rdma_devices
        cls.process_decode = popen_launch_pd_server(
            cls.model,
            cls.decode_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=decode_args,
        )

    def test_gsm8k(self):
        args = SimpleNamespace(
            base_url=self.base_url,
            model=self.model,
            eval_name="gsm8k",
            api="completion",
            max_tokens=512,
            num_examples=200,
            num_threads=128,
        )
        metrics = run_eval(args)
        print(f"{metrics=}")

        self.assertGreater(metrics["score"], 0.24)
        # Wait a little bit so that the memory check happens.
        time.sleep(5)


class TestDisaggregationDecodePPAccuracy(PDDisaggregationServerBase):
    @classmethod
    def setUpClass(cls):
        super().setUpClass()
        cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST)

        # Non blocking start servers
        cls.start_prefill()
        cls.start_decode()

        # Block until both
        cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
        cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)

        cls.launch_lb()

    @classmethod
    def start_prefill(cls):
        prefill_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "prefill",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--pp-size",
            "2",
            "--disable-overlap-schedule",
        ]
        prefill_args += cls.transfer_backend + cls.rdma_devices
        cls.process_prefill = popen_launch_pd_server(
            cls.model,
            cls.prefill_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=prefill_args,
        )

    @classmethod
    def start_decode(cls):
        decode_args = [
            "--trust-remote-code",
            "--disaggregation-mode",
            "decode",
            "--disaggregation-bootstrap-port",
            cls.bootstrap_port,
            "--tp-size",
            "2",
            "--pp-size",
            "2",
            "--base-gpu-id",
            "4",
        ]
        decode_args += cls.transfer_backend + cls.rdma_devices
        cls.process_decode = popen_launch_pd_server(
            cls.model,
            cls.decode_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=decode_args,
        )

    def test_gsm8k(self):
        args = SimpleNamespace(
            base_url=self.base_url,
            model=self.model,
            eval_name="gsm8k",
            api="completion",
            max_tokens=512,
            num_examples=200,
            num_threads=128,
        )
        metrics = run_eval(args)
        print(f"{metrics=}")

        self.assertGreater(metrics["score"], 0.24)
        # Wait a little bit so that the memory check happens.
        time.sleep(5)


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