minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Spec] Retire Spec V1 (#27964)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-06-11 16:15:15 -07:00
committed by GitHub
parent 949326d922
commit c0480a88be
46 changed files with 111 additions and 252 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/dsv4/test_dsv4_flash_mtp_tp8.py
-1
View File
@@ -23,7 +23,6 @@ from sglang.test.test_utils import (
DSV4_FLASH_MODEL_PATH = "sgl-project/DeepSeek-V4-Flash-FP8"
DSV4_FLASH_BASE_ENV = {
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_DSV4_FP4_EXPERTS": "0",
}
test/manual/dsv4/test_dsv4_pro_mtp.py
-1
View File
@@ -40,7 +40,6 @@ HONGLOUMENG_PATH = os.environ.get(
)
DSV4_PRO_BASE_ENV = {
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2": "1",
}
test/registered/8-gpu-models/test_glm_51_fp8.py
-1
View File
@@ -51,7 +51,6 @@ class TestGlm51Fp8(unittest.TestCase):
tp_size=8,
extra_args=COMMON_ARGS + dp_args + MTP_ARGS,
variant="TP8+DP8+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/amd/test_moriep_small.py
+2 -2
View File
@@ -469,6 +469,7 @@ class TestMTPwithTBOLowLatency(CustomTestCase):
"--deepep-mode",
"low_latency",
"--enable-two-batch-overlap",
"--disable-overlap-schedule",
]
)
@@ -476,7 +477,6 @@ class TestMTPwithTBOLowLatency(CustomTestCase):
env["SGLANG_USE_AITER"] = "1"
env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
env["SGLANG_ENABLE_SPEC_V2"] = "false"
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"
@@ -534,6 +534,7 @@ class TestEPLBMoriStat(CustomTestCase):
"normal",
"--expert-distribution-recorder-mode",
"stat",
"--disable-overlap-schedule",
]
)
@@ -541,7 +542,6 @@ class TestEPLBMoriStat(CustomTestCase):
env["SGLANG_USE_AITER"] = "1"
env["SGLANG_MORI_DISPATCH_DTYPE"] = "bf16"
env["SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK"] = "128"
env["SGLANG_ENABLE_SPEC_V2"] = "false"
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
test/registered/ep/test_deepep_large.py
+43 -44
View File
@@ -3,7 +3,6 @@ from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
@@ -87,49 +86,49 @@ class TestDeepseekMTP(CustomTestCase):
def setUpClass(cls):
cls.model = DEFAULT_DEEPEP_MODEL_NAME_FOR_TEST
cls.base_url = DEFAULT_URL_FOR_TEST
with envs.SGLANG_ENABLE_SPEC_V2.override(False):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--tp",
"8",
"--enable-dp-attention",
"--dp",
"8",
"--moe-dense-tp-size",
"1",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--moe-runner-backend",
"deep_gemm",
"--enable-two-batch-overlap",
"--ep-num-redundant-experts",
"32",
"--ep-dispatch-algorithm",
"dynamic",
"--eplb-algorithm",
"deepseek",
"--cuda-graph-bs",
"64", # TODO: increase it to 128 when TBO is supported in draft_extend
"--max-running-requests",
"512",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"1",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"2",
"--disable-radix-cache",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--disable-overlap-schedule",
"--trust-remote-code",
"--tp",
"8",
"--enable-dp-attention",
"--dp",
"8",
"--moe-dense-tp-size",
"1",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--moe-runner-backend",
"deep_gemm",
"--enable-two-batch-overlap",
"--ep-num-redundant-experts",
"32",
"--ep-dispatch-algorithm",
"dynamic",
"--eplb-algorithm",
"deepseek",
"--cuda-graph-bs",
"64", # TODO: increase it to 128 when TBO is supported in draft_extend
"--max-running-requests",
"512",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"1",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"2",
"--disable-radix-cache",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
test/registered/gb300/test_deepseek_v32.py
-1
View File
@@ -59,7 +59,6 @@ class TestDeepseekV32(unittest.TestCase):
]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/gb300/test_deepseek_v32_nvfp4.py
-1
View File
@@ -62,7 +62,6 @@ class TestDeepseekV32Nvfp4(unittest.TestCase):
]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/gb300/test_glm5_fp8.py
-1
View File
@@ -50,7 +50,6 @@ class TestGlm5Fp8(unittest.TestCase):
+ ["--dp-size=4", "--enable-dp-attention"]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/gb300/test_glm5_nvfp4.py
-1
View File
@@ -53,7 +53,6 @@ class TestGlm5Nvfp4(unittest.TestCase):
+ ["--dp-size=4", "--enable-dp-attention"]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/gb300/test_qwen35_fp8.py
-1
View File
@@ -55,7 +55,6 @@ class TestQwen35Fp8(unittest.TestCase):
+ ["--dp-size=4", "--enable-dp-attention"]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/gb300/test_qwen35_nvfp4.py
-1
View File
@@ -59,7 +59,6 @@ class TestQwen35Nvfp4(unittest.TestCase):
+ ["--dp-size=4", "--enable-dp-attention"]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/models_e2e/test_qwen35_fp4_mtp.py
-5
View File
@@ -3,7 +3,6 @@ from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.reasoning_kit import ReasoningTokenUsageMixin
@@ -92,7 +91,6 @@ class TestQwen35FP4MTP(ReasoningTokenUsageMixin, CustomTestCase):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.init_reasoning_token_verifier()
envs.SGLANG_ENABLE_SPEC_V2.set(True)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
@@ -102,7 +100,6 @@ class TestQwen35FP4MTP(ReasoningTokenUsageMixin, CustomTestCase):
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
@@ -117,7 +114,6 @@ class TestQwen35FP4MTPFlashInfer(ReasoningTokenUsageMixin, CustomTestCase):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.init_reasoning_token_verifier()
envs.SGLANG_ENABLE_SPEC_V2.set(True)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
@@ -132,7 +128,6 @@ class TestQwen35FP4MTPFlashInfer(ReasoningTokenUsageMixin, CustomTestCase):
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
test/registered/models_e2e/test_step3p5_flash_chain_mtp.py
-6
View File
@@ -3,7 +3,6 @@ import unittest
import numpy as np
import requests
from sglang.srt.environ import envs
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
@@ -51,11 +50,6 @@ class TestStep3p5FlashChainMTP(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.83
gsm8k_accept_length_thres = 2.6
@classmethod
def setUpClass(cls):
with envs.SGLANG_ENABLE_SPEC_V2.override(True):
super().setUpClass()
def test_logprob_spec_v2_match(self):
"""Verify spec v2 decode logprobs match prefill scoring logprobs.
test/registered/perf/test_dpsk_v3_fp4_4gpu_perf.py
-1
View File
@@ -55,7 +55,6 @@ class TestDeepseekR1FP4Unified(unittest.TestCase):
tp_size=4,
extra_args=base_args + mtp_args,
variant="TP4+MTP",
env={"SGLANG_ENABLE_SPEC_V2": "1"},
),
]
test/registered/sessions/test_streaming_session.py
-1
View File
@@ -57,7 +57,6 @@ class TestStreamingSessionEagleV2RetractLargePage(TestStreamingSession):
"256",
]
env_overrides = [
("SGLANG_ENABLE_SPEC_V2", True),
("SGLANG_TEST_RETRACT", True),
("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", True),
]
test/registered/sessions/test_streaming_session_extra.py
-1
View File
@@ -74,7 +74,6 @@ class TestStreamingSessionEagleV2(StreamingSessionServerBase, StreamingSessionKi
*_EAGLE3_SPEC_ARGS,
]
env_overrides = [
("SGLANG_ENABLE_SPEC_V2", True),
("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", True),
]
test/registered/spec/dflash/test_dflash.py
+5 -3
View File
@@ -28,7 +28,8 @@ class TestDFlashServerBase(CustomTestCase, MatchedStopMixin, GSM8KMixin):
attention_backend = "flashinfer"
page_size = 1
other_launch_args = []
spec_v2 = False
# Base classes exercise the non-overlap (synchronous) scheduling path.
disable_overlap = True
overlap_plan_stream = False
model = DEFAULT_TARGET_MODEL_DFLASH
draft_model = DEFAULT_DRAFT_MODEL_DFLASH
@@ -57,9 +58,10 @@ class TestDFlashServerBase(CustomTestCase, MatchedStopMixin, GSM8KMixin):
"--cuda-graph-bs",
*[str(i) for i in range(1, cls.max_running_requests + 1)],
]
if cls.disable_overlap:
launch_args.append("--disable-overlap-schedule")
launch_args.extend(cls.other_launch_args)
with (
envs.SGLANG_ENABLE_SPEC_V2.override(cls.spec_v2),
envs.SGLANG_ENABLE_OVERLAP_PLAN_STREAM.override(cls.overlap_plan_stream),
envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1),
envs.SGLANG_ENABLE_ASYNC_ASSERT.override(True),
@@ -150,7 +152,7 @@ class TestDFlashServerNoCudaGraph(TestDFlashServerBase):
class TestDFlashServerSpecV2(TestDFlashServerBase):
spec_v2 = True
disable_overlap = False
def test_radix_attention(self):
run_radix_attention_test(self.base_url)
test/registered/spec/eagle/test_eagle_constrained_decoding.py
+11 -10
View File
@@ -1,6 +1,5 @@
import unittest
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.kits.json_constrained_kit import JSONConstrainedMixin
@@ -31,7 +30,8 @@ class TestEagleConstrainedDecoding(
model = DEFAULT_TARGET_MODEL_EAGLE
draft_model = DEFAULT_DRAFT_MODEL_EAGLE
grammar_backend = "xgrammar"
spec_v2 = False
# Run the synchronous (non-overlap) scheduling path.
disable_overlap = True
@classmethod
def setUpClass(cls):
@@ -59,14 +59,15 @@ class TestEagleConstrainedDecoding(
"--grammar-backend",
cls.grammar_backend,
]
if cls.disable_overlap:
launch_args.append("--disable-overlap-schedule")
launch_args.extend(cls.other_launch_args)
with envs.SGLANG_ENABLE_SPEC_V2.override(cls.spec_v2):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=launch_args,
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=launch_args,
)
@classmethod
def tearDownClass(cls):
@@ -74,7 +75,7 @@ class TestEagleConstrainedDecoding(
class TestEagleConstrainedDecodingV2(TestEagleConstrainedDecoding):
spec_v2 = True
disable_overlap = False
if __name__ == "__main__":
test/registered/spec/test_spec_standalone_extra.py
+3 -3
View File
@@ -13,14 +13,14 @@ class TestStandaloneSpeculativeDecodingBase(StandaloneServerBase, CustomTestCase
attention_backend = "fa3"
speculative_eagle_topk = 2
speculative_num_draft_tokens = 7
enable_spec_v2 = False
disable_overlap = True
class TestStandaloneSpeculativeDecodingTriton(StandaloneServerBase, CustomTestCase):
attention_backend = "triton"
speculative_eagle_topk = 2
speculative_num_draft_tokens = 7
enable_spec_v2 = False
disable_overlap = True
enable_deterministic_inference = True
@@ -28,7 +28,7 @@ class TestStandaloneSpeculativeDecodingFlashinfer(StandaloneServerBase, CustomTe
attention_backend = "flashinfer"
speculative_eagle_topk = 2
speculative_num_draft_tokens = 7
enable_spec_v2 = False
disable_overlap = True
if __name__ == "__main__":