397 lines
12 KiB
Python
397 lines
12 KiB
Python
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_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
is_in_ci,
|
|
popen_launch_pd_server,
|
|
)
|
|
|
|
register_cuda_ci(est_time=310, stage="extra-b", runner_config="8-gpu-h200")
|
|
|
|
|
|
@unittest.skipIf(is_in_ci(), "Temporarily disable the flaky test.")
|
|
class TestDisaggregationHybridAttentionGDN(PDDisaggregationServerBase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
super().setUpClass()
|
|
cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct"
|
|
|
|
# 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",
|
|
"4",
|
|
]
|
|
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",
|
|
"4",
|
|
"--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"Evaluation metrics: {metrics}")
|
|
|
|
self.assertGreater(metrics["score"], 0.93)
|
|
|
|
|
|
class TestDisaggregationHybridAttentionGDNExtraBuffer(PDDisaggregationServerBase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
super().setUpClass()
|
|
cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct"
|
|
|
|
# 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",
|
|
"4",
|
|
"--mamba-scheduler-strategy",
|
|
"extra_buffer",
|
|
]
|
|
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",
|
|
"4",
|
|
"--base-gpu-id",
|
|
"4",
|
|
"--mamba-scheduler-strategy",
|
|
"extra_buffer",
|
|
]
|
|
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"Evaluation metrics: {metrics}")
|
|
|
|
# TODO: Fix PD disaggregation accuracy issue (https://github.com/sgl-project/sglang/issues/21744) and increase the threshold back to 0.93.
|
|
self.assertGreater(metrics["score"], 0.90)
|
|
|
|
|
|
class TestDisaggregationHybridAttentionGDNDPDecode(PDDisaggregationServerBase):
|
|
"""Test with prefill tp=2 and decode tp=2/dp=2 with dp-attention enabled."""
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
super().setUpClass()
|
|
cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct"
|
|
|
|
# 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",
|
|
"2",
|
|
]
|
|
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",
|
|
"2",
|
|
"--dp",
|
|
"2",
|
|
"--enable-dp-attention",
|
|
"--enable-dp-lm-head",
|
|
"--base-gpu-id",
|
|
"2",
|
|
]
|
|
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"Evaluation metrics: {metrics}")
|
|
|
|
# TODO: Fix PD disaggregation accuracy issue (https://github.com/sgl-project/sglang/issues/21744) and increase the threshold back to 0.93.
|
|
self.assertGreater(metrics["score"], 0.90)
|
|
|
|
|
|
class TestDisaggregationHybridAttentionMamba(PDDisaggregationServerBase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
super().setUpClass()
|
|
cls.model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
|
|
|
|
# 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",
|
|
"4",
|
|
]
|
|
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",
|
|
"4",
|
|
"--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"Evaluation metrics: {metrics}")
|
|
|
|
self.assertGreater(metrics["score"], 0.87)
|
|
|
|
|
|
class TestDisaggregationHybridAttentionMambaExtraBuffer(PDDisaggregationServerBase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
super().setUpClass()
|
|
cls.model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
|
|
|
|
# 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",
|
|
"4",
|
|
"--mamba-scheduler-strategy",
|
|
"extra_buffer",
|
|
]
|
|
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",
|
|
"4",
|
|
"--base-gpu-id",
|
|
"4",
|
|
"--mamba-scheduler-strategy",
|
|
"extra_buffer",
|
|
]
|
|
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"Evaluation metrics: {metrics}")
|
|
|
|
self.assertGreater(metrics["score"], 0.87)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|