add seed in CPU UTs to avoid flaky failure (#7333)

test/srt/cpu/test_activation.py

@@ -8,6 +8,8 @@ from utils import SiluAndMul, precision
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestActivation(CustomTestCase):
     M = [128, 129, 257]

test/srt/cpu/test_decode.py

@@ -6,6 +6,8 @@ from torch.nn.functional import scaled_dot_product_attention
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestDecodeAttention(CustomTestCase):
     def _run_sdpa_forward_decode(

test/srt/cpu/test_extend.py

@@ -6,6 +6,8 @@ from torch.nn.functional import scaled_dot_product_attention
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestExtendAttention(CustomTestCase):
 

test/srt/cpu/test_gemm.py

@@ -14,6 +14,8 @@ from utils import (
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class Mod(nn.Module):
     def __init__(self, input_channel, output_channel, has_bias):

test/srt/cpu/test_mla.py

@@ -8,6 +8,8 @@ from utils import precision
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestMLA(CustomTestCase):
     def _run_sdpa_forward_decode(

test/srt/cpu/test_moe.py

@@ -8,6 +8,8 @@ import torch
 
 kernel = torch.ops.sgl_kernel
 
+torch.manual_seed(0)
+
 from utils import (
     BLOCK_K,
     BLOCK_N,

test/srt/cpu/test_norm.py

@@ -8,6 +8,8 @@ from utils import make_non_contiguous, precision
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestNorm(CustomTestCase):
     M = [4096, 1024]

test/srt/cpu/test_rope.py

@@ -10,6 +10,8 @@ from sglang.srt.layers.rotary_embedding import (
 )
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestROPE(CustomTestCase):
     def test_deepseek_v2_rope(self):

test/srt/cpu/test_shared_expert.py

@@ -22,6 +22,8 @@ from utils import (
 
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 class TestSharedExpert(CustomTestCase):
     M = [2, 121]

test/srt/cpu/test_topk.py

@@ -13,6 +13,8 @@ from sglang.srt.layers.moe.topk import grouped_topk_gpu as native_grouped_topk
 from sglang.srt.models.llama4 import Llama4MoE
 from sglang.test.test_utils import CustomTestCase
 
+torch.manual_seed(0)
+
 
 # This is used by the Deepseek-V2 model
 class TestGroupedTopK(CustomTestCase):