chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Tiny apply gsm8k mixin to ngram test (#15606)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-12-24 01:30:26 +08:00
committed by GitHub
parent 5f3a47d8a7
commit bd572360f3
6 changed files with 31 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/models/test_ministral3_models.py
+1 -1
View File
@@ -9,7 +9,7 @@ MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
accuracy = 0.6
gsm8k_accuracy_thres = 0.6
model = MODEL
other_args = ["--trust-remote-code"]
test/srt/models/test_nvidia_nemotron_nano_v2.py
+5 -5
View File
@@ -7,19 +7,19 @@ from sglang.test.server_fixtures.default_fixture import DefaultServerBase
class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256"]
class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8"
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skipIf(not is_blackwell(), "NVFP4 only supported on blackwell")
class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
accuracy = 0.855
gsm8k_accuracy_thres = 0.855
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-NVFP4"
other_args = ["--max-mamba-cache-size", "256"]
@@ -29,7 +29,7 @@ class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
@@ -60,7 +60,7 @@ class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase)
class TestNvidiaNemotronNanoV2SpeculativeDecodingBF16Cache(
GSM8KMixin, DefaultServerBase
):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
test/srt/models/test_nvidia_nemotron_nano_v2_vl.py
+1 -1
View File
@@ -9,7 +9,7 @@ MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, DefaultServerBase):
accuracy = 0.87
gsm8k_accuracy_thres = 0.87
model = MODEL
other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]