minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Spec v2 tree drafting (topk>1) with page_size>1 (#26972)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-06-06 12:00:27 -07:00
committed by GitHub
parent 88a7b0fd30
commit 84ca0ffb8c
8 changed files with 228 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/spec/eagle/test_spec_eagle_page.py
+4 -11
View File
@@ -1,7 +1,7 @@
"""page_size > 1 variants (flashinfer).
"""page_size > 1 variants at topk=1 (flashinfer).
EAGLE3 page64 (spec v2) + EAGLE/Llama-2 page4 (topk1 and topk8, spec v1).
Runs on the cheap (5090) runner.
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.
"""
import unittest
@@ -15,7 +15,7 @@ from sglang.test.kits.spec_server_kits import (
)
from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
register_cuda_ci(est_time=540, stage="base-b", runner_config="1-gpu-small")
register_cuda_ci(est_time=360, stage="base-b", runner_config="1-gpu-small")
class TestEagle3Page64(Eagle3Base, SpecAccuracyKit, SpecLogprobKit, SpecFeatureKit):
@@ -35,12 +35,5 @@ class TestEagleLlama2Page4Topk1(EagleLlama2Base, SpecAccuracyKit, SpecFeatureKit
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_topk_page.py
+41
View File
@@ -0,0 +1,41 @@
"""topk > 1 tree drafting at page_size > 1 (EAGLE3 topk8 + EAGLE/Llama-2 topk8).
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.)
"""
import unittest
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,
SpecFeatureKit,
)
from sglang.test.server_fixtures.spec_eagle_fixture import Eagle3Base, EagleLlama2Base
register_cuda_ci(est_time=720, stage="base-b", runner_config="1-gpu-small")
class TestEagle3Page64Topk8(Eagle3Base, SpecAccuracyKit, SpecFeatureKit):
"""EAGLE3 topk=8 tree + page_size=64 (spec v2)."""
page_size = 64
spec_topk = 8
spec_tokens = 32
disable_overlap = False
cuda_graph_max_bs = 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()