/test/registered/amd/test_moriep_small.py

import os
import unittest
from types import SimpleNamespace

import requests

from sglang.srt.server_args import ZMQ_TCP_PORT_DELTA
from sglang.srt.utils import kill_process_tree
from sglang.srt.utils.network import is_port_available
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
    DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST,
    DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST_NEXTN,
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    CustomTestCase,
    popen_launch_server,
)


def wait_all_ports_release(base_url, timeout_s=60):
    """Wait until all derived ports are fully released."""
    import time

    port = int(base_url.split(":")[-1])

    # See https://github.com/sgl-project/sglang/blob/495ef8ec64b6b937e59cd530ad3150172061a008/python/sglang/srt/server_args.py#L6958-L6969
    offsets = [
        0,  # no offset
        ZMQ_TCP_PORT_DELTA,  # dist_init_port
        ZMQ_TCP_PORT_DELTA + 1,  # detokenizer_port
        ZMQ_TCP_PORT_DELTA + 2,  # rpc_port
        ZMQ_TCP_PORT_DELTA + 3,  # metrics_port
        ZMQ_TCP_PORT_DELTA + 4,  # scheduler_input_port
    ]
    for _ in range(timeout_s):
        if all(is_port_available(port + off) for off in offsets):
            return
        time.sleep(1)
    # Best-effort: log but don't raise so tearDown doesn't break the next class.
    print(f"Warning: some ports still occupied after {timeout_s}s")


register_amd_ci(est_time=5400, suite="stage-c-test-large-8-gpu-amd")

common_args = [
    "--tp-size",
    "8",
    "--ep-size",
    "8",
    "--dp-size",
    "8",
    "--enable-dp-attention",
    "--moe-a2a-backend",
    "mori",
    "--trust-remote-code",
    "--load-balance-method",
    "round_robin",
    "--moe-dense-tp-size",
    "1",
    "--enable-dp-lm-head",
    "--mem-fraction-static",
    "0.7",  # relax for mi300x
    "--chunked-prefill-size",
    "1024",
    "--max-running-requests",
    "128",
    "--context-length",
    "4096",
    "--max-total-tokens",
    "32768",
    "--attention-backend",
    "aiter",
    "--cuda-graph-max-bs-decode",
    "32",
]

eplb_args = [
    "--enable-eplb",
    "--ep-num-redundant-experts",
    "32",
    "--eplb-rebalance-num-iterations",
    "50",
    "--expert-distribution-recorder-buffer-size",
    "50",
    "--ep-dispatch-algorithm",
    "static",
]

mtp_args = [
    "--speculative-algo",
    "EAGLE",
    "--speculative-num-steps",
    "3",
    "--speculative-eagle-topk",
    "1",
    "--speculative-num-draft-tokens",
    "4",
    "--speculative-draft-model-path",
    DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST_NEXTN,
]


class TestPureDP(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")

        self.assertGreaterEqual(metrics["accuracy"], 0.90)


class TestMTP(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args + mtp_args

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")
        self.assertGreaterEqual(metrics["accuracy"], 0.90)

        server_info = requests.get(self.base_url + "/server_info")
        avg_spec_accept_length = server_info.json()["internal_states"][0][
            "avg_spec_accept_length"
        ]
        print(f"{avg_spec_accept_length=}")
        self.assertGreaterEqual(avg_spec_accept_length, 2.8)


class TestNormal(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args + [
            "--deepep-mode",
            "normal",
        ]

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")

        self.assertGreaterEqual(metrics["accuracy"], 0.90)


class TestLowLatency(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args + [
            "--deepep-mode",
            "low_latency",
        ]

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory
        # FIXME(billishyahao): enable p2p due to no rdma devices on CI machine
        # env["MORI_DISABLE_P2P"] = "1"

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")

        self.assertGreaterEqual(metrics["accuracy"], 0.90)


class TestTBOwithNormal(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args + [
            "--deepep-mode",
            "normal",
            "--enable-two-batch-overlap",
        ]

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")

        self.assertGreaterEqual(metrics["accuracy"], 0.90)


class TestTBOwithLowLatency(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = common_args + [
            "--deepep-mode",
            "low_latency",
            "--enable-two-batch-overlap",
        ]

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory
        # FIXME(billishyahao): enable p2p due to no rdma devices on CI machine
        # env["MORI_DISABLE_P2P"] = "1"

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")

        self.assertGreaterEqual(metrics["accuracy"], 0.90)


class TestMTPwithTBONormal(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = (
            common_args
            + mtp_args
            + [
                "--deepep-mode",
                "normal",
                "--enable-two-batch-overlap",
            ]
        )

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory
        env["MORI_ENABLE_SDMA"] = "true"

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")
        self.assertGreaterEqual(metrics["accuracy"], 0.90)

        server_info = requests.get(self.base_url + "/server_info")
        avg_spec_accept_length = server_info.json()["internal_states"][0][
            "avg_spec_accept_length"
        ]
        print(f"{avg_spec_accept_length=}")
        self.assertGreaterEqual(avg_spec_accept_length, 2.8)


class TestMTPwithTBOLowLatency(CustomTestCase):

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = (
            common_args
            + mtp_args
            + [
                "--deepep-mode",
                "low_latency",
                "--enable-two-batch-overlap",
                "--disable-overlap-schedule",
            ]
        )

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory
        # FIXME(billishyahao): enable p2p due to no rdma devices on CI machine
        # env["MORI_DISABLE_P2P"] = "1"
        env["MORI_ENABLE_SDMA"] = "true"

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")
        self.assertGreaterEqual(metrics["accuracy"], 0.90)

        server_info = requests.get(self.base_url + "/server_info")
        avg_spec_accept_length = server_info.json()["internal_states"][0][
            "avg_spec_accept_length"
        ]
        print(f"{avg_spec_accept_length=}")
        self.assertGreaterEqual(avg_spec_accept_length, 2.8)


class TestEPLBMoriStat(CustomTestCase):
    """EPLB with mori backend, stat mode (on_select_experts path)."""

    @classmethod
    def setUpClass(cls):
        cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = (
            common_args
            + eplb_args
            + [
                "--deepep-mode",
                "normal",
                "--expert-distribution-recorder-mode",
                "stat",
                "--disable-overlap-schedule",
            ]
        )

        env = dict(os.environ)
        env["SGLANG_USE_AITER"] = "1"
        env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
        env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
        env["SGLANG_EPLB_ROCM_P2P_BATCH_CHUNK_SIZE"] = "32"
        env["MORI_SHMEM_MODE"] = "ISOLATION"  # avoid out of symmetric heap memory
        # FIXME(billishyahao): enable p2p due to no rdma devices on CI machine
        # env["MORI_DISABLE_P2P"] = "1"

        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 5,
            other_args=other_args,
            env=env,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)
        wait_all_ports_release(cls.base_url)

    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(f"{metrics=}")
        self.assertGreaterEqual(metrics["accuracy"], 0.9)


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