/test/registered/quant/test_fp8_blockwise_gemm.py

import unittest
from types import SimpleNamespace
from urllib.parse import urlparse

from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    popen_launch_server,
    try_cached_model,
)

register_cuda_ci(est_time=430, stage="extra-b", runner_config="4-gpu-b200")

MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
MXFP8_MODEL_PATH = "zianglih/Qwen3-4B-Instruct-2507-MXFP8"


class FP8BlockwiseGemmBase:
    backend = None

    @classmethod
    def setUpClass(cls):
        if cls.backend is None:
            raise NotImplementedError("Subclass must set 'backend' attribute")
        cls.model = try_cached_model(MODEL_PATH)
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--fp8-gemm-backend",
            cls.backend,
        ]
        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):
        parsed_url = urlparse(self.base_url)
        args = SimpleNamespace(
            base_url=self.base_url,
            model=self.model,
            eval_name="gsm8k",
            api="completion",
            max_tokens=512,
            num_examples=1319,
            num_threads=200,
            num_shots=8,
        )
        metrics = run_eval(args)
        print(metrics)

        self.assertGreaterEqual(metrics["score"], 0.8)


class MXFP8GemmBase:
    backend = None

    @classmethod
    def setUpClass(cls):
        if cls.backend is None:
            raise NotImplementedError("Subclass must set 'backend' attribute")
        cls.model = try_cached_model(MXFP8_MODEL_PATH)
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--fp8-gemm-backend",
            cls.backend,
            # TODO: pin tc_piecewise — default `breakable` prefill runs MXFP8 RMSNorm in bf16, hurting accuracy; unrelated to BCG-default change.
            "--cuda-graph-backend-prefill",
            "tc_piecewise",
        ]
        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):
        parsed_url = urlparse(self.base_url)
        args = SimpleNamespace(
            base_url=self.base_url,
            model=self.model,
            eval_name="gsm8k",
            api="completion",
            max_tokens=512,
            num_examples=1319,
            num_threads=200,
            num_shots=8,
        )
        metrics = run_eval(args)
        print(metrics)

        self.assertGreaterEqual(metrics["score"], 0.8)


class TestFP8BlockwiseGemmTriton(FP8BlockwiseGemmBase, unittest.TestCase):
    backend = "triton"


class TestFP8BlockwiseGemmDeepGemm(FP8BlockwiseGemmBase, unittest.TestCase):
    backend = "deep_gemm"


@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestFP8BlockwiseGemmFlashinferTrtllm(FP8BlockwiseGemmBase, unittest.TestCase):
    backend = "flashinfer_trtllm"


@unittest.skipIf(get_device_sm() != 90, "Test requires CUDA SM 90")
class TestFP8BlockwiseGemmFlashinferDeepGemm(FP8BlockwiseGemmBase, unittest.TestCase):
    backend = "flashinfer_deepgemm"


@unittest.skip("Currently PCG capture takes too long to complete, disable until fixed")
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestMXFP8GemmTriton(MXFP8GemmBase, unittest.TestCase):
    backend = "triton"


@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestMXFP8GemmFlashinferTrtllm(MXFP8GemmBase, unittest.TestCase):
    backend = "flashinfer_trtllm"


@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestMXFP8GemmFlashinferCutlass(MXFP8GemmBase, unittest.TestCase):
    backend = "flashinfer_cutlass"


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