Files
sglang/test/registered/disaggregation/test_disaggregation_different_tp.py
T

599 lines
18 KiB
Python

import os
import unittest
from types import SimpleNamespace
from sglang.srt.environ import envs
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_HYBRID_GDN_SMALL_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
popen_launch_pd_server,
try_cached_model,
)
register_cuda_ci(est_time=375, stage="base-c", runner_config="8-gpu-h20")
class TestDisaggregationMooncakePrefillLargerTP(PDDisaggregationServerBase):
@classmethod
def setUpClass(cls):
super().setUpClass()
# Temporarily disable JIT DeepGEMM
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST_MLA)
# 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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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",
"--base-gpu-id",
"4",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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"], 0.60)
class TestDisaggregationMooncakeDecodeLargerTP(PDDisaggregationServerBase):
@classmethod
def setUpClass(cls):
super().setUpClass()
# Temporarily disable JIT DeepGEMM
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST_MLA)
# 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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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"], 0.60)
class TestDisaggregationMooncakeMHAPrefillLargerTP(PDDisaggregationServerBase):
@classmethod
def setUpClass(cls):
super().setUpClass()
# Temporarily disable JIT DeepGEMM
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
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",
"4",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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",
"--base-gpu-id",
"4",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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"], 0.60)
class TestDisaggregationMooncakeMHADecodeLargerTP(PDDisaggregationServerBase):
@classmethod
def setUpClass(cls):
super().setUpClass()
# Temporarily disable JIT DeepGEMM
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
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",
"2",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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"], 0.60)
STAGING_ENV = {
"SGLANG_DISAGG_STAGING_BUFFER": "1",
"SGLANG_DISAGG_STAGING_BUFFER_SIZE_MB": "64",
"SGLANG_DISAGG_STAGING_POOL_SIZE_MB": "1024",
}
class TestDisaggregationStagingPrefillLargerTP(PDDisaggregationServerBase):
"""Prefill TP=4 -> Decode TP=2 with staging buffer enabled (MHA model)."""
@classmethod
def setUpClass(cls):
super().setUpClass()
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST)
cls.start_prefill()
cls.start_decode()
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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
env = {**os.environ, **STAGING_ENV}
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=prefill_args,
env=env,
)
@classmethod
def start_decode(cls):
decode_args = [
"--trust-remote-code",
"--disaggregation-mode",
"decode",
"--disaggregation-bootstrap-port",
cls.bootstrap_port,
"--tp",
"2",
"--base-gpu-id",
"4",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
decode_args += cls.transfer_backend + cls.rdma_devices
env = {**os.environ, **STAGING_ENV}
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=decode_args,
env=env,
)
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"[Staging PrefillLargerTP] Evaluation metrics: {metrics}")
self.assertGreater(metrics["score"], 0.60)
class TestDisaggregationStagingDecodeLargerTP(PDDisaggregationServerBase):
"""Prefill TP=2 -> Decode TP=4 with staging buffer enabled (MHA model)."""
@classmethod
def setUpClass(cls):
super().setUpClass()
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST)
cls.start_prefill()
cls.start_decode()
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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
env = {**os.environ, **STAGING_ENV}
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=prefill_args,
env=env,
)
@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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
decode_args += cls.transfer_backend + cls.rdma_devices
env = {**os.environ, **STAGING_ENV}
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=decode_args,
env=env,
)
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"[Staging DecodeLargerTP] Evaluation metrics: {metrics}")
self.assertGreater(metrics["score"], 0.60)
class TestDisaggregationGDNHybridHeteroTP(PDDisaggregationServerBase):
"""Prefill TP=1 -> Decode TP=4 on a GDN-hybrid (gated-delta-net) model.
Exercises the heterogeneous attn-TP *scatter* path (prefill attn_tp <
decode attn_tp), where two independent bugs corrupt accuracy without the
fix in this change set:
1. GDN conv_state is cat([query|key|value]) with each sub-block
independently head-sharded; a single contiguous slice straddles the
q/k/v boundaries and delivers wrong channels.
2. GQA KV heads are replicated when num_key_value_heads < decode attn_tp;
the scatter head map must use integer division (tp_rank //
num_kv_head_replicas), not modulo.
Without the fix gsm8k collapses (~0.4); with it, it recovers to agg level.
"""
@classmethod
def setUpClass(cls):
super().setUpClass()
envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
cls.model = try_cached_model(DEFAULT_HYBRID_GDN_SMALL_MODEL_NAME_FOR_TEST)
cls.start_prefill()
cls.start_decode()
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",
"1",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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",
"--enable-metrics",
"--enable-request-time-stats-logging",
]
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"[GDNHybridHeteroTP] Evaluation metrics: {metrics}")
self.assertGreater(metrics["score"], 0.60)
if __name__ == "__main__":
unittest.main()