EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[CI] improve disaggregation CI. (#11264)

Browse Source 
Signed-off-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Liangsheng Yin
2025-10-08 21:40:56 +08:00
committed by GitHub
parent e3bb7f5ae6
commit c882b5ae75
9 changed files with 107 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
test/srt/hicache/test_disaggregation_hicache.py
View File

@@ -70,11 +70,8 @@ class DisaggregationHiCacheBase(TestDisaggregationBase):
"wait_complete",
"--mem-fraction-static",
"0.8",
"--disaggregation-ib-device",
"mlx5_roce0",
"--disaggregation-transfer-backend",
"mooncake",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
env = {
**os.environ,
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.temp_dir,
@@ -148,11 +145,8 @@ class TestDisaggregationPrefillWithHiCache(DisaggregationHiCacheBase):
"0.8",
"--base-gpu-id",
"1",
"--disaggregation-ib-device",
"mlx5_roce0",
"--disaggregation-transfer-backend",
"mooncake",
]
decode_args += cls.transfer_backend + cls.rdma_devices
env = {
**os.environ,
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.temp_dir,
@@ -201,10 +195,6 @@ class TestDisaggregationDecodeWithHiCache(DisaggregationHiCacheBase):
"0.8",
"--base-gpu-id",
"1",
"--disaggregation-ib-device",
"mlx5_roce0",
"--disaggregation-transfer-backend",
"mooncake",
"--disaggregation-decode-enable-offload-kvcache",
"--hicache-ratio",
"1.2",
@@ -215,6 +205,7 @@ class TestDisaggregationDecodeWithHiCache(DisaggregationHiCacheBase):
"--hicache-storage-prefetch-policy",
"wait_complete",
]
decode_args += cls.transfer_backend + cls.rdma_devices
env = {
**os.environ,
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.temp_dir,

5
test/srt/hicache/test_hicache_storage_mooncake_backend.py
View File

@@ -15,6 +15,7 @@ import requests
from test_hicache_storage_file_backend import HiCacheStorageBaseMixin
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_disaggregation_utils import get_rdma_devices_args
from sglang.test.test_utils import (
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
CustomTestCase,
@@ -192,7 +193,7 @@ class HiCacheStorageMooncakeBackendBaseMixin(HiCacheStorageBaseMixin):
"""Get additional server arguments specific to configuration - override in subclasses"""
server_args = {
"--tp-size": 1,
"--tp-size": 2,
"--hicache-ratio": 2,
"--hicache-storage-backend": "mooncake",
}
@@ -202,7 +203,7 @@ class HiCacheStorageMooncakeBackendBaseMixin(HiCacheStorageBaseMixin):
"MOONCAKE_MASTER": f"127.0.0.1:{cls.mooncake_master_port}",
"MOONCAKE_PROTOCOL": "rdma",
"MC_MS_AUTO_DISC": "0",
"MOONCAKE_DEVICE": "mlx5_roce0,mlx5_roce1",
"MOONCAKE_DEVICE": get_rdma_devices_args(),
"MOONCAKE_TE_META_DATA_SERVER": f"http://127.0.0.1:{cls.mooncake_metadata_port}/metadata",
"MOONCAKE_GLOBAL_SEGMENT_SIZE": "4294967296", # 4 GiB
}

4
test/srt/run_suite.py
View File

@@ -134,11 +134,13 @@ suites = {
TestFile("lora/test_lora_tp.py", 116),
TestFile("rl/test_update_weights_from_distributed.py", 103),
TestFile("test_data_parallelism.py", 73),
TestFile("test_disaggregation.py", 499),
TestFile("test_dp_attention.py", 594),
TestFile("test_load_weights_from_remote_instance.py", 72),
TestFile("test_patch_torch.py", 19),
TestFile("test_release_memory_occupation.py", 257),
TestFile("hicache/test_hicache_storage_file_backend.py", 200),
TestFile("hicache/test_hicache_storage_mooncake_backend.py", 400),
TestFile("hicache/test_hicache_storage_3fs_backend.py", 200),
],
"per-commit-4-gpu": [
@@ -149,9 +151,7 @@ suites = {
TestFile("test_multi_instance_release_memory_occupation.py", 64),
],
"per-commit-8-gpu": [
TestFile("hicache/test_hicache_storage_mooncake_backend.py", 400),
TestFile("lora/test_lora_llama4.py", 400),
TestFile("test_disaggregation.py", 499),
TestFile("test_disaggregation_dp_attention.py", 155),
TestFile("test_disaggregation_different_tp.py", 600),
TestFile("test_disaggregation_pp.py", 140),

48
test/srt/test_disaggregation.py
View File

@@ -40,10 +40,9 @@ class TestDisaggregationAccuracy(TestDisaggregationBase):
"--disaggregation-mode",
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
"1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -58,12 +57,11 @@ class TestDisaggregationAccuracy(TestDisaggregationBase):
"--disaggregation-mode",
"decode",
"--tp",
"2",
"1",
"--base-gpu-id",
"2",
"--disaggregation-ib-device",
"mlx5_roce2,mlx5_roce3",
"1",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -171,10 +169,9 @@ class TestDisaggregationMooncakeFailure(TestDisaggregationBase):
"--disaggregation-mode",
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
"1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -189,12 +186,11 @@ class TestDisaggregationMooncakeFailure(TestDisaggregationBase):
"--disaggregation-mode",
"decode",
"--tp",
"2",
"1",
"--base-gpu-id",
"2",
"--disaggregation-ib-device",
"mlx5_roce2,mlx5_roce3",
"1",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -270,10 +266,9 @@ class TestDisaggregationMooncakeSpec(TestDisaggregationBase):
"--disaggregation-mode",
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
"1",
] + cls.spec_args
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -288,12 +283,11 @@ class TestDisaggregationMooncakeSpec(TestDisaggregationBase):
"--disaggregation-mode",
"decode",
"--tp",
"2",
"1",
"--base-gpu-id",
"2",
"--disaggregation-ib-device",
"mlx5_roce2,mlx5_roce3",
"1",
] + cls.spec_args
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -346,10 +340,9 @@ class TestDisaggregationSimulatedRetract(TestDisaggregationBase):
"--disaggregation-mode",
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
"1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -364,12 +357,11 @@ class TestDisaggregationSimulatedRetract(TestDisaggregationBase):
"--disaggregation-mode",
"decode",
"--tp",
"2",
"1",
"--base-gpu-id",
"2",
"--disaggregation-ib-device",
"mlx5_roce2,mlx5_roce3",
"1",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,

24
test/srt/test_disaggregation_different_tp.py
View File

@@ -41,9 +41,8 @@ class TestDisaggregationMooncakePrefillLargerTP(TestDisaggregationBase):
"prefill",
"--tp",
"4",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -61,9 +60,8 @@ class TestDisaggregationMooncakePrefillLargerTP(TestDisaggregationBase):
"2",
"--base-gpu-id",
"4",
"--disaggregation-ib-device",
"mlx5_roce4,mlx5_roce5",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -115,9 +113,8 @@ class TestDisaggregationMooncakeDecodeLargerTP(TestDisaggregationBase):
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -135,9 +132,8 @@ class TestDisaggregationMooncakeDecodeLargerTP(TestDisaggregationBase):
"4",
"--base-gpu-id",
"4",
"--disaggregation-ib-device",
"mlx5_roce4,mlx5_roce5",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -189,9 +185,8 @@ class TestDisaggregationMooncakeMHAPrefillLargerTP(TestDisaggregationBase):
"prefill",
"--tp",
"4",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -209,9 +204,8 @@ class TestDisaggregationMooncakeMHAPrefillLargerTP(TestDisaggregationBase):
"2",
"--base-gpu-id",
"4",
"--disaggregation-ib-device",
"mlx5_roce4,mlx5_roce5",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
@@ -263,9 +257,8 @@ class TestDisaggregationMooncakeMHADecodeLargerTP(TestDisaggregationBase):
"prefill",
"--tp",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -283,9 +276,8 @@ class TestDisaggregationMooncakeMHADecodeLargerTP(TestDisaggregationBase):
"4",
"--base-gpu-id",
"4",
"--disaggregation-ib-device",
"mlx5_roce4,mlx5_roce5",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,

6
test/srt/test_disaggregation_dp_attention.py
View File

@@ -45,9 +45,8 @@ class TestDisaggregationDPAttention(TestDisaggregationBase):
"--dp",
"2",
"--enable-dp-attention",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -68,9 +67,8 @@ class TestDisaggregationDPAttention(TestDisaggregationBase):
"--enable-dp-attention",
"--base-gpu-id",
"2",
"--disaggregation-ib-device",
"mlx5_roce2,mlx5_roce3",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,

6
test/srt/test_disaggregation_pp.py
View File

@@ -37,10 +37,9 @@ class TestDisaggregationPPAccuracy(TestDisaggregationBase):
"2",
"--pp-size",
"2",
"--disaggregation-ib-device",
"mlx5_roce0,mlx5_roce1",
"--disable-overlap-schedule",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
@@ -58,9 +57,8 @@ class TestDisaggregationPPAccuracy(TestDisaggregationBase):
"2",
"--base-gpu-id",
"4",
"--disaggregation-ib-device",
"mlx5_roce4,mlx5_roce5",
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,