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

Tiny cleanup the models' name in test_utils (#15920)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-12-27 14:13:23 +08:00
committed by GitHub
parent 29ce7b3612
commit 9ad546d7e8
15 changed files with 88 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/server_fixtures/eagle_fixture.py
+4 -4
View File
@@ -6,8 +6,8 @@ import requests
from sglang.srt.utils.common import kill_process_tree
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -24,8 +24,8 @@ PROMPTS = [
class EagleServerBase(CustomTestCase):
target_model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
target_model = DEFAULT_TARGET_MODEL_EAGLE
draft_model = DEFAULT_DRAFT_MODEL_EAGLE
spec_algo = "EAGLE"
spec_steps = 5
spec_topk = 8
python/sglang/test/test_utils.py
+18 -12
View File
@@ -91,18 +91,24 @@ DEFAULT_MODEL_NAME_FOR_TEST_AWQ_INT4 = (
"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"
)
# EAGLE
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST = "meta-llama/Llama-2-7b-chat-hf"
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST = "lmsys/sglang-EAGLE-llama2-chat-7B"
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3 = "meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_EAGLE_DP_ATTENTION_TARGET_MODEL_FOR_TEST = "Qwen/Qwen3-30B-A3B"
DEFAULT_EAGLE_DP_ATTENTION_DRAFT_MODEL_FOR_TEST = "Tengyunw/qwen3_30b_moe_eagle3"
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3 = "lmsys/sglang-EAGLE3-LLaMA3.1-Instruct-8B"
DEFAULT_STANDALONE_SPECULATIVE_TARGET_MODEL_FOR_TEST = (
"meta-llama/Llama-3.1-8B-Instruct"
)
DEFAULT_STANDALONE_SPECULATIVE_DRAFT_MODEL_FOR_TEST = "meta-llama/Llama-3.2-1B-Instruct"
DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST = "Qwen/Qwen2.5-Coder-7B-Instruct"
# EAGLE2 algorithm models
DEFAULT_TARGET_MODEL_EAGLE = "meta-llama/Llama-2-7b-chat-hf"
DEFAULT_DRAFT_MODEL_EAGLE = "lmsys/sglang-EAGLE-llama2-chat-7B"
# EAGLE3 model
DEFAULT_TARGET_MODEL_EAGLE3 = "meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_DRAFT_MODEL_EAGLE3 = "lmsys/sglang-EAGLE3-LLaMA3.1-Instruct-8B"
# EAGLE2 with DP-Attention models
DEFAULT_TARGET_MODEL_EAGLE_DP_ATTN = "Qwen/Qwen3-30B-A3B"
DEFAULT_DRAFT_MODEL_EAGLE_DP_ATTN = "Tengyunw/qwen3_30b_moe_eagle3"
# Standalone speculative decoding models
DEFAULT_TARGET_MODEL_STANDALONE = "meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_DRAFT_MODEL_STANDALONE = "meta-llama/Llama-3.2-1B-Instruct"
# N-gram speculative decoding models
DEFAULT_TARGET_MODEL_NGRAM = "Qwen/Qwen2.5-Coder-7B-Instruct"
# Other use cases
DEFAULT_AUTOROUND_MODEL_NAME_FOR_TEST = (
test/registered/attention/test_fa3.py
+5 -7
View File
@@ -8,8 +8,8 @@ from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE3,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -30,7 +30,7 @@ OFFLINE_MODE = False
# Change the path below when OFFLINE_MODE is True.
OFFLINE_PATH_DICT = {
DEFAULT_MODEL_NAME_FOR_TEST: "/shared/public/elr-models/meta-llama/Meta-Llama-3.1-8B-Instruct",
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3: "/shared/public/elr-models/jamesliu1/sglang-EAGLE3-Llama-3.1-Instruct-8B",
DEFAULT_DRAFT_MODEL_EAGLE3: "/shared/public/elr-models/jamesliu1/sglang-EAGLE3-Llama-3.1-Instruct-8B",
DEFAULT_MODEL_NAME_FOR_TEST_MLA: "/shared/public/sharing/deepseek/dsv3-test/snapshots/",
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN: "/shared/public/sharing/deepseek/dsv3-test-NextN/snapshots/",
GSM_DATASET_PATH: "/shared/public/data/gsm8k/test.jsonl",
@@ -39,9 +39,7 @@ OFFLINE_PATH_DICT = {
if OFFLINE_MODE:
DEFAULT_MODEL_NAME_FOR_TEST = OFFLINE_PATH_DICT[DEFAULT_MODEL_NAME_FOR_TEST]
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3 = OFFLINE_PATH_DICT[
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3
]
DEFAULT_DRAFT_MODEL_EAGLE3 = OFFLINE_PATH_DICT[DEFAULT_DRAFT_MODEL_EAGLE3]
DEFAULT_MODEL_NAME_FOR_TEST_MLA = OFFLINE_PATH_DICT[DEFAULT_MODEL_NAME_FOR_TEST_MLA]
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN = OFFLINE_PATH_DICT[
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN
@@ -154,7 +152,7 @@ class TestFlashAttention3SpeculativeDecode(BaseFlashAttentionTest):
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model-path",
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_DRAFT_MODEL_EAGLE3,
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
@@ -188,7 +186,7 @@ class TestFlashAttention3SpeculativeDecodeTopk(BaseFlashAttentionTest):
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model-path",
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_DRAFT_MODEL_EAGLE3,
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
test/registered/attention/test_hybrid_attn_backend.py
+5 -5
View File
@@ -8,10 +8,10 @@ from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -59,7 +59,7 @@ class TestHybridAttnBackendBase(CustomTestCase):
envs.SGLANG_ENABLE_JIT_DEEPGEMM.override(False),
):
if cls.speculative_decode:
model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
model = DEFAULT_TARGET_MODEL_EAGLE
else:
model = cls.model
cls.process = popen_launch_server(
@@ -129,7 +129,7 @@ class TestHybridAttnBackendSpeculativeDecodingPrefillBackend(TestHybridAttnBacke
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
@@ -152,7 +152,7 @@ class TestHybridAttnBackendSpeculativeDecodingDecodeBackend(TestHybridAttnBacken
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
test/registered/spec/eagle/test_eagle_constrained_decoding.py
+4 -4
View File
@@ -6,8 +6,8 @@ from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.json_constrained_kit import TestJSONConstrainedMixin
from sglang.test.kits.regex_constrained_kit import TestRegexConstrainedMixin
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -27,8 +27,8 @@ class TestEagleConstrainedDecoding(
spec_draft_tokens = 6
page_size = 1
other_launch_args = []
model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
model = DEFAULT_TARGET_MODEL_EAGLE
draft_model = DEFAULT_DRAFT_MODEL_EAGLE
grammar_backend = "xgrammar"
eagle_v2 = False
test/registered/spec/eagle/test_eagle_infer_a.py
+17 -17
View File
@@ -10,11 +10,11 @@ from sglang.srt.utils import kill_process_tree
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_DRAFT_MODEL_EAGLE3,
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TARGET_MODEL_EAGLE3,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -31,8 +31,8 @@ decode_tolerance: float = 5e-2
class TestEAGLEEngine(CustomTestCase):
BASE_CONFIG = {
"model_path": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
"speculative_draft_model_path": DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
"model_path": DEFAULT_TARGET_MODEL_EAGLE,
"speculative_draft_model_path": DEFAULT_DRAFT_MODEL_EAGLE,
"speculative_algorithm": "EAGLE",
"speculative_num_steps": 5,
"speculative_eagle_topk": 4,
@@ -133,7 +133,7 @@ class TestEAGLEEngine(CustomTestCase):
"skip_special_tokens": False,
}
tokenizer = get_tokenizer(DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST)
tokenizer = get_tokenizer(DEFAULT_TARGET_MODEL_EAGLE)
output = engine.generate(prompt, params)["text"]
print(f"{output=}")
@@ -187,8 +187,8 @@ class TestEAGLEEngineTokenMap(TestEAGLEEngine):
class TestEAGLE3Engine(TestEAGLEEngine):
BASE_CONFIG = {
"model_path": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
"speculative_draft_model_path": DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
"model_path": DEFAULT_TARGET_MODEL_EAGLE3,
"speculative_draft_model_path": DEFAULT_DRAFT_MODEL_EAGLE3,
"speculative_algorithm": "EAGLE3",
"speculative_num_steps": 5,
"speculative_eagle_topk": 16,
@@ -206,8 +206,8 @@ class TestEAGLE3Engine(TestEAGLEEngine):
class TestEAGLERadixCache(CustomTestCase):
BASE_CONFIG = {
"model_path": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
"speculative_draft_model_path": DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
"model_path": DEFAULT_TARGET_MODEL_EAGLE3,
"speculative_draft_model_path": DEFAULT_DRAFT_MODEL_EAGLE3,
"speculative_algorithm": "EAGLE3",
"speculative_num_steps": 2,
"speculative_eagle_topk": 2,
@@ -309,14 +309,14 @@ class TestEAGLEDraftExtend(CustomTestCase):
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
1,
"--speculative-eagle-topk",
@@ -375,14 +375,14 @@ class TestEAGLEDraftExtendFlashinfer(TestEAGLEDraftExtend):
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
1,
"--speculative-eagle-topk",
@@ -404,14 +404,14 @@ class TestEAGLEDraftExtendTriton(TestEAGLEDraftExtend):
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
1,
"--speculative-eagle-topk",
test/registered/spec/eagle/test_eagle_infer_b.py
+2 -5
View File
@@ -15,10 +15,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_gsm8k_eval
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.server_fixtures.eagle_fixture import EagleServerBase
from sglang.test.test_utils import (
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
run_logprob_check,
)
from sglang.test.test_utils import DEFAULT_TARGET_MODEL_EAGLE, run_logprob_check
register_cuda_ci(est_time=1100, suite="stage-b-test-small-1-gpu")
@@ -262,7 +259,7 @@ class TestEAGLEServerBasic(EagleServerBase):
response = requests.post(
self.base_url + "/v1/chat/completions",
json={
"model": DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
"model": DEFAULT_TARGET_MODEL_EAGLE,
"messages": messages,
"temperature": 0,
"response_format": {"type": "json_object"},
test/registered/spec/eagle/test_eagle_infer_beta.py
+4 -4
View File
@@ -8,8 +8,8 @@ from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kits.matched_stop_kit import MatchedStopMixin
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -27,8 +27,8 @@ class TestEagleServerBase(CustomTestCase, MatchedStopMixin):
spec_draft_tokens = 6
page_size = 1
other_launch_args = []
model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
model = DEFAULT_TARGET_MODEL_EAGLE
draft_model = DEFAULT_DRAFT_MODEL_EAGLE
@classmethod
def setUpClass(cls):
test/registered/spec/test_ngram_speculative_decoding.py
+2 -2
View File
@@ -5,7 +5,7 @@ from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.test_utils import (
DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST,
DEFAULT_TARGET_MODEL_NGRAM,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -32,7 +32,7 @@ DEFAULT_SERVER_ARGS = [
class TestNgramSpeculativeDecodingBase(GSM8KMixin, CustomTestCase):
model = DEFAULT_NGRAM_SPECULATIVE_TARGET_MODEL_FOR_TEST
model = DEFAULT_TARGET_MODEL_NGRAM
base_url = DEFAULT_URL_FOR_TEST
gsm8k_accuracy_thres = 0.79 # derived tests need to override this
gsm8k_accept_length_thres = 1.8 # derived spec decoding tests need to override this
test/srt/hicache/test_hicache_variants.py
+4 -4
View File
@@ -12,10 +12,10 @@ from sglang.bench_serving import get_tokenizer
from sglang.srt.utils import is_hip, kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3,
DEFAULT_DRAFT_MODEL_EAGLE3,
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_TARGET_MODEL_EAGLE3,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -113,7 +113,7 @@ class TestHiCacheMLA(HiCacheBaseServer, HiCacheEvalMixin, HiCacheMGSMEvalMixin):
class TestHiCacheEagle(HiCacheBaseServer, HiCacheEvalMixin):
"""HiCache with EAGLE speculative decoding tests"""
model_name = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3
model_name = DEFAULT_TARGET_MODEL_EAGLE3
needs_tokenizer = True
hicache_args = [
"--enable-hierarchical-cache",
@@ -124,7 +124,7 @@ class TestHiCacheEagle(HiCacheBaseServer, HiCacheEvalMixin):
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model-path",
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3,
DEFAULT_DRAFT_MODEL_EAGLE3,
"--speculative-num-steps",
2,
"--speculative-eagle-topk",
test/srt/openai_server/features/test_openai_server_hidden_states.py
+6 -6
View File
@@ -6,9 +6,9 @@ import openai
from sglang.srt.utils import kill_process_tree
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -259,10 +259,10 @@ class TestOpenAIServerWithEAGLEAndHiddenStatesEnabled(
):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
cls.model = DEFAULT_TARGET_MODEL_EAGLE
cls.base_url = DEFAULT_URL_FOR_TEST
cls.api_key = "sk-123456"
cls.speculative_draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
cls.speculative_draft_model = DEFAULT_DRAFT_MODEL_EAGLE
cls.speculative_algorithm = "EAGLE"
cls.process = popen_launch_server(
cls.model,
@@ -272,7 +272,7 @@ class TestOpenAIServerWithEAGLEAndHiddenStatesEnabled(
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
5,
"--speculative-eagle-topk",
@@ -289,7 +289,7 @@ class TestOpenAIServerWithEAGLEAndHiddenStatesEnabled(
],
)
cls.base_url += "/v1"
cls.tokenizer = get_tokenizer(DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST)
cls.tokenizer = get_tokenizer(DEFAULT_TARGET_MODEL_EAGLE)
cls.return_hidden_states = [False, True]
cls.use_list_input = [True, False]
cls.parallel_sample_nums = [1]
test/srt/test_bench_serving.py
+4 -4
View File
@@ -5,14 +5,14 @@ import unittest
import requests
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_FP8,
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
@@ -328,7 +328,7 @@ class TestBenchServing(CustomTestCase):
def test_online_latency_eagle(self):
res = run_bench_serving(
model=DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
model=DEFAULT_TARGET_MODEL_EAGLE,
num_prompts=300,
request_rate=8,
sharegpt_context_len=3072,
@@ -338,7 +338,7 @@ class TestBenchServing(CustomTestCase):
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
test/srt/test_disaggregation_basic.py
+4 -4
View File
@@ -12,9 +12,9 @@ from sglang.test.server_fixtures.disaggregation_fixture import (
PDDisaggregationServerBase,
)
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
popen_launch_pd_server,
)
@@ -280,8 +280,8 @@ class TestDisaggregationMooncakeSpec(PDDisaggregationServerBase):
@classmethod
def setUpClass(cls):
super().setUpClass()
cls.model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
cls.draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
cls.model = DEFAULT_TARGET_MODEL_EAGLE
cls.draft_model = DEFAULT_DRAFT_MODEL_EAGLE
cls.spec_args = [
"--speculative-algorithm",
"EAGLE",
test/srt/test_eagle_dp_attention.py
+4 -4
View File
@@ -6,8 +6,8 @@ import requests
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_EAGLE_DP_ATTENTION_DRAFT_MODEL_FOR_TEST,
DEFAULT_EAGLE_DP_ATTENTION_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE_DP_ATTN,
DEFAULT_TARGET_MODEL_EAGLE_DP_ATTN,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
class TestEAGLE3EngineDPAttention(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_EAGLE_DP_ATTENTION_TARGET_MODEL_FOR_TEST
cls.model = DEFAULT_TARGET_MODEL_EAGLE_DP_ATTN
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
@@ -35,7 +35,7 @@ class TestEAGLE3EngineDPAttention(CustomTestCase):
"--speculative-num-draft-tokens",
"32",
"--speculative-draft-model-path",
DEFAULT_EAGLE_DP_ATTENTION_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_EAGLE_DP_ATTN,
"--tp-size",
"2",
"--dp-size",
test/srt/test_standalone_speculative_decoding.py
+5 -5
View File
@@ -7,8 +7,8 @@ from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_STANDALONE_SPECULATIVE_DRAFT_MODEL_FOR_TEST,
DEFAULT_STANDALONE_SPECULATIVE_TARGET_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_STANDALONE,
DEFAULT_TARGET_MODEL_STANDALONE,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -26,7 +26,7 @@ DEFAULT_SERVER_ARGS = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-draft-model-path",
DEFAULT_STANDALONE_SPECULATIVE_DRAFT_MODEL_FOR_TEST,
DEFAULT_DRAFT_MODEL_STANDALONE,
"--speculative-num-steps",
"4",
"--speculative-eagle-topk",
@@ -40,8 +40,8 @@ DEFAULT_SERVER_ARGS = [
class TestStandaloneSpeculativeDecodingBase(CustomTestCase):
model = DEFAULT_STANDALONE_SPECULATIVE_TARGET_MODEL_FOR_TEST
draft_model = DEFAULT_STANDALONE_SPECULATIVE_DRAFT_MODEL_FOR_TEST
model = DEFAULT_TARGET_MODEL_STANDALONE
draft_model = DEFAULT_DRAFT_MODEL_STANDALONE
base_url = DEFAULT_URL_FOR_TEST
accuracy_threshold = 0.7 # derived tests need to override this
spec_decode_threshold = 3.6 # derived spec decoding tests need to override this