[CI] Collapse the EAGLE launch matrix and the scoring engine boots on the per-commit runners (#33756)

test/registered/spec/dflash/test_dflash.py

@@ -161,7 +161,7 @@ class TestDFlashServerNoCudaGraph(TestDFlashServerBase):
     other_launch_args = ["--disable-cuda-graph"]
 
 
-class TestDFlashServerSpecV2(TestDFlashServerBase):
+class TestDFlashServerOverlap(TestDFlashServerBase):
     disable_overlap = False
 
     def test_radix_attention(self):
@@ -169,7 +169,7 @@ class TestDFlashServerSpecV2(TestDFlashServerBase):
         assert self.process.poll() is None
 
 
-class TestDFlashServerSpecV2PlanStream(TestDFlashServerSpecV2):
+class TestDFlashServerOverlapPlanStream(TestDFlashServerOverlap):
     overlap_plan_stream = True
 
 

test/registered/spec/eagle/test_spec_eagle.py

@@ -1,5 +1,7 @@
-"""EAGLE3 spec-decoding core: overlap (spec v2) x no-overlap (spec v1) matrix,
-same standard config (topk=1, page_size=1), only ``disable_overlap`` differs.
+"""EAGLE3 spec-decoding core: overlap x no-overlap matrix at the standard
+config (topk=1, page_size=1); only ``disable_overlap`` differs. Both run the
+same EAGLEWorkerV2 -- the scheduler just drives it synchronously when overlap
+is off.
 flashinfer is pinned (the 5090 default) so a default-selection change can't
 silently alter what this exercises.
 """
@@ -35,13 +37,13 @@ class _Core(Eagle3Base):
 
 
 class TestEagle3Overlap(_Core, *_KITS):
-    """Spec v2 (overlap scheduler on)."""
+    """Overlap scheduler on."""
 
     disable_overlap = False
 
 
 class TestEagle3NoOverlap(_Core, *_KITS):
-    """Spec v1 (overlap scheduler off)."""
+    """Overlap scheduler off (synchronous)."""
 
     disable_overlap = True
 

test/registered/spec/eagle/test_spec_eagle_fa3.py

@@ -10,7 +10,6 @@ from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
-    SpecCorrectnessKit,
     SpecFeatureKit,
     SpecLogprobKit,
     SpecPenaltyKit,
@@ -18,11 +17,16 @@ from sglang.test.kits.spec_server_kits import (
 )
 from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
 
-register_cuda_ci(est_time=600, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=250, stage="base-b", runner_config="1-gpu-large")
 
 
-class TestEagle3Fa3(Eagle3Base, SpecCorrectnessKit, SpecAccuracyKit, SpecLogprobKit):
-    """EAGLE3 spec v2 topk=1 on fa3 (the H200 default backend)."""
+class TestEagle3Fa3(Eagle3Base, SpecAccuracyKit, SpecLogprobKit):
+    """EAGLE3 topk=1 on fa3 (the H200 default backend), overlap on.
+
+    No SpecCorrectnessKit: those checks are scheduler/sampling behaviour, which
+    the 5090 runs already cover. Logprob losslessness stays -- it reads through
+    the verify output, which the attention unit cases do not reach.
+    """
 
     attention_backend = "fa3"
     disable_overlap = False
@@ -37,7 +41,7 @@ class TestEagleLlama2Fa3Page256(
     SpecPerfKit,
     SpecFeatureKit,
 ):
-    """EAGLE/Llama-2 topk=5 tree on fa3 + page_size=256 (spec v1)."""
+    """EAGLE/Llama-2 topk=5 tree on fa3 + page_size=256, overlap off."""
 
     spec_topk = 5
     spec_steps = 8

test/registered/spec/eagle/test_spec_eagle_page.py

@@ -1,7 +1,7 @@
-"""page_size > 1 variants at topk=1 (flashinfer).
+"""EAGLE3 chain drafting (topk=1) at page_size > 1, flashinfer.
 
-EAGLE3 page64 (spec v2) + EAGLE/Llama-2 page4 (spec v1). topk>1 page variants
-live in test_spec_eagle_topk.py. Runs on the cheap (5090) runner.
+topk=1 takes its own fast path in the draft worker, so this cell is not
+covered by the tree variants in test_spec_eagle_topk_page.py.
 """
 
 import unittest
@@ -10,30 +10,28 @@ from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
+    SpecCorrectnessKit,
     SpecFeatureKit,
     SpecLogprobKit,
 )
-from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
+from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base
 
-register_cuda_ci(est_time=360, stage="base-b", runner_config="1-gpu-small")
+register_cuda_ci(est_time=230, stage="base-b", runner_config="1-gpu-small")
 
 
-class TestEagle3Page64(Eagle3Base, SpecAccuracyKit, SpecLogprobKit, SpecFeatureKit):
-    """EAGLE3 spec v2, page_size=64 (flashinfer): + logprob losslessness."""
+class TestEagle3Page64(
+    Eagle3Base,
+    SpecCorrectnessKit,
+    SpecAccuracyKit,
+    SpecLogprobKit,
+    SpecFeatureKit,
+):
+    """Overlap scheduler, page_size=64: + logprob losslessness."""
 
     page_size = 64
     disable_overlap = False
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 
-class TestEagleLlama2Page4Topk1(EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit):
-    """Llama-2 topk=1 + page_size=4."""
-
-    spec_topk = 1
-    spec_tokens = 6
-    page_size = 4
-    env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/spec/eagle/test_spec_eagle_parity.py

@@ -27,7 +27,7 @@ class _Eagle3ParityBase(Eagle3Base):
 
 @unittest.skipIf(_is_xpu, "CUDA runner only")
 class TestEagle3ParityCUDA(SpecParityKit, _Eagle3ParityBase):
-    """EAGLE3 spec v2 (flashinfer, overlap) greedy output == non-spec reference.
+    """EAGLE3 (flashinfer, overlap) greedy output == non-spec reference.
 
     SpecParityKit is first so its setUpClass runs the reference server (and tears
     it down) before the fixture launches the spec server -- sequential, one model

test/registered/spec/eagle/test_spec_eagle_stress.py

@@ -1,18 +1,15 @@
-"""Perf + stress: throughput, retract-under-pressure, abort storms, timeouts.
+"""Perf + stress: throughput and retract-under-pressure.
 
-These need memory headroom / measure load behavior, so they run on the large
-(Hopper) runner.
+These need memory headroom / measure load behaviour, so they run on the large
+(Hopper) runner. The scheduler timeout paths carry no spec-specific state, so
+they live in unit/managers/test_scheduler_timeouts.py plus the cheap e2e in
+scheduler/test_scheduler_control.py.
 """
 
 import unittest
 
 from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
-from sglang.test.kits.abort_timeout_kit import (
-    AbortAllMixin,
-    RunningTimeoutTwoWaveMixin,
-    WaitingTimeoutMixin,
-)
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
     SpecFeatureKit,
@@ -20,11 +17,11 @@ from sglang.test.kits.spec_server_kits import (
 )
 from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
 
-register_cuda_ci(est_time=780, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=440, stage="base-b", runner_config="1-gpu-large")
 
 
 class TestEagle3Perf(Eagle3Base, SpecPerfKit):
-    """Decode throughput (max_new_tokens=1) on EAGLE3 spec v2."""
+    """Decode throughput (max_new_tokens=1) on EAGLE3."""
 
     disable_overlap = False
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
@@ -42,7 +39,7 @@ class TestEagleLlama2Retract(EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit):
 
 
 class TestEagle3Topk16V2Retract(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
-    """EAGLE3 topk=16 tree on spec v2 under retract; must not leak KV. Stresses
+    """EAGLE3 topk=16 tree under retract; must not leak KV. Stresses
     the accepted-path KV move (move_accept_tokens_to_target_kvcache)."""
 
     spec_topk = 16
@@ -58,27 +55,5 @@ class TestEagle3Topk16V2Retract(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
     )
 
 
-class TestEagleLlama2AbortAll(EagleLlama2Base, AbortAllMixin):
-    abort_all_max_new_tokens = 4000
-    env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
-
-
-class TestEagleLlama2WaitingTimeout(EagleLlama2Base, WaitingTimeoutMixin):
-    max_running_requests = 1
-    env_overrides = (
-        (envs.SGLANG_REQ_WAITING_TIMEOUT, 0.001),
-        (envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),
-    )
-
-
-class TestEagleLlama2RunningTimeout(EagleLlama2Base, RunningTimeoutTwoWaveMixin):
-    # Regression: https://github.com/sgl-project/sglang/pull/18760
-    max_running_requests = 16
-    env_overrides = (
-        (envs.SGLANG_REQ_RUNNING_TIMEOUT, 3),
-        (envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),
-    )
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/spec/eagle/test_spec_eagle_topk.py

@@ -1,8 +1,7 @@
-"""topk > 1 tree drafting (EAGLE3 topk16 + EAGLE/Llama-2 topk8).
+"""topk > 1 tree drafting at page_size=1 (EAGLE3 topk16 + EAGLE/Llama-2 topk8).
 
-topk > 1 routes to spec v1, except page_size==1 which can also stay on spec v2
-(overlap). flashinfer is pinned because this runs on the cheap (5090) runner,
-where fa3 (Hopper-only) isn't available -- functional sanity only, no perf/stress.
+flashinfer is pinned because this runs on the cheap (5090) runner, where fa3
+(Hopper-only) isn't available -- functional sanity only, no perf/stress.
 (topk > 1 on fa3 is covered on the Hopper runner in test_spec_eagle_fa3.py.)
 """
 
@@ -10,24 +9,37 @@ import unittest
 
 from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.abort_timeout_kit import AbortAllMixin
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
     SpecCorrectnessKit,
     SpecFeatureKit,
+    SpecHiddenStatesKit,
     SpecLogprobKit,
     SpecPenaltyKit,
 )
 from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
 
-register_cuda_ci(est_time=1180, stage="base-b", runner_config="1-gpu-small")
+register_cuda_ci(est_time=870, stage="base-b", runner_config="1-gpu-small")
 
 
-class TestEagle3Topk16(Eagle3Base, SpecCorrectnessKit, SpecAccuracyKit, SpecLogprobKit):
-    """EAGLE3 topk=16 tree (spec v1): correctness + gsm8k + logprob losslessness."""
+class TestEagle3Topk16(
+    Eagle3Base,
+    SpecCorrectnessKit,
+    SpecAccuracyKit,
+    SpecLogprobKit,
+    SpecFeatureKit,
+    SpecHiddenStatesKit,
+):
+    """EAGLE3 topk=16 tree, overlap scheduler: guards the accepted-path
+    compaction (via logprob_decode_match_prefill) and the per-request
+    hidden-state stride slicing that the same compaction feeds.
+    """
 
     spec_topk = 16
     spec_tokens = 64
-    disable_overlap = True  # synchronous baseline; SpecV2 subclass flips overlap on
+    disable_overlap = False
+    enable_return_hidden_states = True
     cuda_graph_max_bs_decode = 5
     acc_length_thres = 3.1
     batch_accept_len_thres = 1.75
@@ -35,13 +47,6 @@ class TestEagle3Topk16(Eagle3Base, SpecCorrectnessKit, SpecAccuracyKit, SpecLogp
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 
-class TestEagle3Topk16SpecV2(TestEagle3Topk16, SpecFeatureKit):
-    """EAGLE3 topk=16 tree on spec v2 (overlap, page1): guards the v2 tree path's
-    accepted-path compaction, validated by logprob_spec_v2_match."""
-
-    disable_overlap = False
-
-
 class TestEagleLlama2Suite(
     EagleLlama2Base,
     SpecCorrectnessKit,
@@ -49,9 +54,16 @@ class TestEagleLlama2Suite(
     SpecLogprobKit,
     SpecPenaltyKit,
     SpecFeatureKit,
+    AbortAllMixin,
 ):
-    """EAGLE/Llama-2 topk=8 full coverage (kits listed in bases)."""
+    """EAGLE/Llama-2 topk=8 full coverage (kits listed in bases).
 
+    Hosts AbortAllMixin: aborting mid-decode has to release the tree draft
+    state, and the strict mem check below turns a leak into a failure. It needs
+    no server flags of its own, so it rides this launch.
+    """
+
+    abort_all_max_new_tokens = 4000
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 

test/registered/spec/eagle/test_spec_eagle_topk_page.py

@@ -1,9 +1,10 @@
-"""topk > 1 tree drafting at page_size > 1 (EAGLE3 topk8 + EAGLE/Llama-2 topk8).
+"""EAGLE3 tree drafting (topk > 1) at page_size > 1, flashinfer (fa3 is
+Hopper-only).
 
-page64 stays on spec v2 (overlap), page4 runs on spec v1 (no overlap). flashinfer is
-pinned because this runs on the cheap (5090) runner, where fa3 (Hopper-only) isn't
-available -- functional sanity only, no perf/stress. (page>1 topk>1 on fa3 is covered
-on the Hopper runner in test_spec_eagle_fa3.py.)
+page_size=4 with 32 draft tokens spreads the draft window over several pages --
+the layout the unit fixture refuses to build (tree draft is pinned to
+page_size=1 there, see speculative_draft_runner.py). The window-inside-one-page
+regime is covered by test_spec_eagle_fa3.py page256 on the Hopper runner.
 """
 
 import unittest
@@ -13,29 +14,26 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
     SpecFeatureKit,
+    SpecLogprobKit,
 )
-from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
+from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base
 
-register_cuda_ci(est_time=720, stage="base-b", runner_config="1-gpu-small")
+register_cuda_ci(est_time=345, stage="base-b", runner_config="1-gpu-small")
 
 
-class TestEagle3Page64Topk8(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
-    """EAGLE3 topk=8 tree + page_size=64 (spec v2)."""
+class TestEagle3Page4Topk8(Eagle3Base, SpecAccuracyKit, SpecLogprobKit, SpecFeatureKit):
+    """Overlap scheduler, topk=8 tree, page_size=4."""
 
-    page_size = 64
+    page_size = 4
     spec_topk = 8
     spec_tokens = 32
     disable_overlap = False
+    # Preset accept-length values are topk=1 numbers -- loose for a topk=8
+    # tree; tighten once CI reports the actuals.
+    gsm8k_accept_len_thres = 2.0
     cuda_graph_max_bs_decode = 5
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 
-class TestEagleLlama2Page4Topk8(EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit):
-    """Llama-2 topk>1 tree + page_size=4 (spec v1)."""
-
-    page_size = 4
-    env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/spec/eagle/test_spec_eagle_triton.py

@@ -1,6 +1,8 @@
-"""triton attention backend (EAGLE3 topk=1 chain + EAGLE/Llama-2 topk=8 tree).
+"""triton attention backend, EAGLE3 chain drafting.
 
-triton runs everywhere, so this stays on the cheap (5090) runner.
+triton runs everywhere, so this stays on the cheap (5090) runner. triton tree
+verify is covered by attention/unittests/dense/test_triton.py, and the tree
+accept-path compaction e2e lives in test_spec_eagle_topk.py.
 """
 
 import unittest
@@ -11,13 +13,12 @@ from sglang.test.kits.matched_stop_kit import MatchedStopMixin
 from sglang.test.kits.spec_server_kits import (
     SpecAccuracyKit,
     SpecFeatureKit,
-    SpecHiddenStatesKit,
     SpecLogprobKit,
     SpecPenaltyKit,
 )
-from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
+from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base
 
-register_cuda_ci(est_time=350, stage="base-b", runner_config="1-gpu-small")
+register_cuda_ci(est_time=230, stage="base-b", runner_config="1-gpu-small")
 
 
 class TestEagle3Triton(
@@ -28,7 +29,7 @@ class TestEagle3Triton(
     SpecPenaltyKit,
     SpecFeatureKit,
 ):
-    """EAGLE3 spec v2 on triton (kits listed in bases)."""
+    """Overlap scheduler on triton (kits listed in bases)."""
 
     attention_backend = "triton"
     max_running_requests = 64
@@ -38,19 +39,5 @@ class TestEagle3Triton(
     env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
 
 
-class TestEagleLlama2Triton(
-    EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit, SpecHiddenStatesKit
-):
-    """EAGLE/Llama-2 topk=8 tree on triton.
-
-    Hosts SpecHiddenStatesKit: topk>1 exercises the tree accept-path
-    compaction that the per-req hidden-state stride slicing depends on.
-    """
-
-    attention_backend = "triton"
-    enable_return_hidden_states = True
-    env_overrides = ((envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY, 1),)
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/spec/test_spec_standalone.py

@@ -4,24 +4,29 @@ from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.json_constrained_kit import JSONConstrainedMixin
 from sglang.test.kits.regex_constrained_kit import RegexConstrainedMixin
 from sglang.test.server_fixtures.standalone_fixture import StandaloneServerBase
-from sglang.test.test_utils import CustomTestCase
+from sglang.test.test_utils import CustomTestCase, is_in_ci
 
-# V2 standalone speculative decoding tests (FA3, Triton, FlashInfer backends).
+# V2 standalone speculative decoding. CI runs only fa3 (the backend this is
+# deployed on); triton / flashinfer stay runnable locally, and their spec verify
+# numerics live in attention/unittests/dense/test_{triton,flashinfer}.py.
 # Non-V2 backends moved to test_spec_standalone_extra.py.
-register_cuda_ci(est_time=450, stage="base-b", runner_config="1-gpu-large")
+register_cuda_ci(est_time=170, stage="base-b", runner_config="1-gpu-large")
 
 
-class TestStandaloneV2SpeculativeDecodingBase(StandaloneServerBase, CustomTestCase):
+class TestStandaloneV2SpeculativeDecodingBase(
+    StandaloneServerBase, CustomTestCase, RegexConstrainedMixin, JSONConstrainedMixin
+):
+    # Hosts the constrained mixins: overlap is on, so they exercise the
+    # grammar barrier path.
     attention_backend = "fa3"
 
 
-class TestStandaloneV2SpeculativeDecodingTriton(
-    StandaloneServerBase, CustomTestCase, RegexConstrainedMixin, JSONConstrainedMixin
-):
-    # Constrained mixins reuse this server; overlap on -> grammar barrier path.
+@unittest.skipIf(is_in_ci(), "CI covers fa3 only; run locally for triton.")
+class TestStandaloneV2SpeculativeDecodingTriton(StandaloneServerBase, CustomTestCase):
     attention_backend = "triton"
 
 
+@unittest.skipIf(is_in_ci(), "CI covers fa3 only; run locally for flashinfer.")
 class TestStandaloneV2SpeculativeDecodingFlashinfer(
     StandaloneServerBase, CustomTestCase
 ):