[Platform] Deprecate seed_everything (#31659)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

tests/kernels/core/test_pos_encoding.py

@@ -9,7 +9,7 @@ import torch
 
 from tests.kernels.allclose_default import get_default_atol, get_default_rtol
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 IS_NEOX_STYLE = [True, False]
 DTYPES = [torch.bfloat16, torch.float]
@@ -79,7 +79,7 @@ def test_rotary_embedding(
     if rotary_dim is None:
         rotary_dim = head_size
 
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
     torch.set_default_device(device)
     if rotary_dim is None:
         rotary_dim = head_size

tests/kernels/mamba/test_causal_conv1d.py

@@ -12,7 +12,7 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
     causal_conv1d_fn,
     causal_conv1d_update,
 )
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 
 def causal_conv1d_ref(
@@ -154,7 +154,7 @@ def test_causal_conv1d_update(dim, width, seqlen, has_bias, silu_activation, ity
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch = 2
     x = torch.randn(batch, dim, seqlen, device=device, dtype=itype)
     x_ref = x.clone()
@@ -201,7 +201,7 @@ def test_causal_conv1d_update_with_batch_gather(
         rtol, atol = 1e-2, 5e-2
 
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
 
     padding = 5 if with_padding else 0
     padded_batch_size = batch_size + padding
@@ -278,7 +278,7 @@ def test_causal_conv1d_varlen(
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     seqlens = []
     batch_size = batch
     padding = 3 if with_padding else 0

tests/kernels/mamba/test_mamba_mixer2.py

@@ -12,8 +12,8 @@ from vllm.distributed.parallel_state import (
     initialize_model_parallel,
 )
 from vllm.model_executor.layers.mamba.mamba_mixer2 import Mixer2RMSNormGated
-from vllm.platforms import current_platform
 from vllm.utils.system_utils import update_environment_variables
+from vllm.utils.torch_utils import set_random_seed
 
 
 @multi_gpu_test(num_gpus=2)
@@ -68,7 +68,7 @@ def mixer2_gated_norm_tensor_parallel(
     dtype: torch.dtype,
     device: str,
 ):
-    current_platform.seed_everything(0)
+    set_random_seed(0)
 
     device = torch.device(f"cuda:{local_rank}")
     torch.cuda.set_device(device)

tests/kernels/mamba/test_mamba_ssm.py

@@ -13,7 +13,7 @@ from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
     selective_scan_fn,
     selective_state_update,
 )
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 
 def selective_state_update_ref(
@@ -271,7 +271,7 @@ def test_selective_scan(
         rtolw = max(rtolw, rtol)
         atolw = max(atolw, atol)
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch_size = 1
     dim = 4
     dstate = 8
@@ -401,7 +401,7 @@ def test_selective_state_update(dim, dstate, has_z, itype):
         if torch.version.hip:
             atol *= 2
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch_size = 1
     state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device)
     x = torch.randn(batch_size, dim, device=device, dtype=itype)
@@ -438,7 +438,7 @@ def test_selective_state_update_varlen(dim, dstate, has_z, itype, max_seq_len):
         if torch.version.hip:
             atol *= 2
     # set seed
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch_size = 4
     token_counts = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
     total_tokens = int(token_counts.sum().item())
@@ -857,7 +857,7 @@ def test_selective_state_update_with_num_accepted_tokens(
         if torch.version.hip:
             atol *= 2
 
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch_size = 4
 
     tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
@@ -983,7 +983,7 @@ def test_selective_state_update_varlen_with_num_accepted(
         if torch.version.hip:
             atol *= 2
 
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     batch_size = 4
 
     tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)

tests/kernels/mamba/test_mamba_ssm_ssd.py

@@ -9,7 +9,7 @@ from einops import rearrange, repeat
 from vllm.model_executor.layers.mamba.ops.ssd_combined import (
     mamba_chunk_scan_combined_varlen,
 )
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.attention.backends.mamba2_attn import compute_varlen_chunk_metadata
 
 # Added by the IBM Team, 2024
@@ -82,7 +82,7 @@ def ssd_minimal_discrete(X, A, B, C, block_len, initial_states=None):
 
 
 def generate_random_inputs(batch_size, seqlen, n_heads, d_head, itype, device="cuda"):
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     A = -torch.exp(torch.rand(n_heads, dtype=itype, device=device))
     dt = F.softplus(
         torch.randn(batch_size, seqlen, n_heads, dtype=itype, device=device) - 4

tests/kernels/moe/modular_kernel_tools/make_feature_matrix.py

@@ -10,7 +10,7 @@ from tqdm import tqdm
 
 from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.fused_moe.config import FUSED_MOE_UNQUANTIZED_CONFIG
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 from .common import (
     Config,
@@ -40,7 +40,7 @@ def rank_worker(
     config: Config,
     weights: WeightTensors,
 ):
-    current_platform.seed_everything(pgi.rank)
+    set_random_seed(pgi.rank)
 
     # sanity check
     from vllm import envs

tests/kernels/moe/modular_kernel_tools/profile_modular_kernel.py

@@ -9,7 +9,7 @@ from typing import Any
 import torch
 
 from vllm.config import VllmConfig
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 from .common import Config, RankTensors, WeightTensors, make_modular_kernel
 from .parallel_utils import ProcessGroupInfo, parallel_launch_with_config
@@ -82,7 +82,7 @@ def rank_worker(
     config: Config,
     weights: WeightTensors,
 ):
-    current_platform.seed_everything(pgi.rank)
+    set_random_seed(pgi.rank)
 
     # sanity check
     from vllm import envs

tests/kernels/moe/test_batched_moe.py

@@ -21,6 +21,7 @@ from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl
+from vllm.utils.torch_utils import set_random_seed
 
 MNK_FACTORS = [
     (1, 128, 128),
@@ -115,7 +116,7 @@ def test_batched_mm(
 ):
     """Note: float8_e4m3fn is not supported on CUDA architecture < 89,
     and those tests will be skipped on unsupported hardware."""
-    current_platform.seed_everything(7)
+    set_random_seed(7)
 
     use_fp8_w8a8 = dtype == torch.float8_e4m3fn
 
@@ -252,7 +253,7 @@ def test_fused_moe_batched_experts(
 ):
     """Note: float8_e4m3fn is not supported on CUDA architecture < 89,
     and those tests will be skipped on unsupported hardware."""
-    current_platform.seed_everything(7)
+    set_random_seed(7)
 
     use_fp8_w8a8 = dtype == torch.float8_e4m3fn
 

tests/kernels/moe/test_cpu_fused_moe.py

@@ -8,6 +8,7 @@ from tests.kernels.allclose_default import get_default_atol, get_default_rtol
 from vllm._custom_ops import cpu_fused_moe, cpu_prepack_moe_weight
 from vllm.model_executor.layers.activation import SiluAndMul, SwigluOAIAndMul
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 if not current_platform.is_cpu():
     pytest.skip("skipping CPU-only tests", allow_module_level=True)
@@ -114,7 +115,7 @@ def test_cpu_fused_moe(
     act: str,
     isa: str,
 ):
-    current_platform.seed_everything(0)
+    set_random_seed(0)
 
     topk_num = max(expert_num // 2, 1)
     up_dim = 2 * intermediate_size

tests/kernels/moe/test_cutlass_moe.py

@@ -20,6 +20,7 @@ from vllm.model_executor.layers.fused_moe.cutlass_moe import (
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 NUM_EXPERTS = [40, 64]
 TOP_KS = [6, 8]
@@ -277,7 +278,7 @@ def test_cutlass_moe_8_bit_no_graph(
     workspace_init,
     ep_size: int | None = None,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token, per_out_ch)
@@ -332,7 +333,7 @@ def test_cutlass_moe_8_bit_cuda_graph(
     monkeypatch,
     workspace_init,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         dtype = torch.half
@@ -469,7 +470,7 @@ def test_run_cutlass_moe_fp8(
     ep_size: int,
     workspace_init,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     with set_current_vllm_config(vllm_config):
         mt = MOETensors8Bit.make_moe_tensors_8bit(
             m, k, n, e, per_act_token, per_out_channel

tests/kernels/moe/test_deepep_deepgemm_moe.py

@@ -22,13 +22,13 @@ from vllm.model_executor.layers.fused_moe.config import (
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
 from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
-from vllm.platforms import current_platform
 from vllm.utils.deep_gemm import (
     get_mk_alignment_for_contiguous_layout,
     is_deep_gemm_e8m0_used,
     is_deep_gemm_supported,
 )
 from vllm.utils.import_utils import has_deep_ep, has_deep_gemm
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.worker.workspace import init_workspace_manager
 
 from ...utils import multi_gpu_test
@@ -367,7 +367,7 @@ def _test_deepep_deepgemm_moe(
     device = torch.device(f"cuda:{pgi.local_rank}")
     init_workspace_manager(device)
 
-    current_platform.seed_everything(pgi.rank)
+    set_random_seed(pgi.rank)
 
     w1 = w1.to(device=torch.cuda.current_device())
     w2 = w2.to(device=torch.cuda.current_device())
@@ -456,7 +456,7 @@ def test_ht_deepep_deepgemm_moe(
     """
 
     m, n, k = mnk
-    current_platform.seed_everything(7)
+    set_random_seed(7)
 
     if topk > num_experts:
         pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
@@ -531,7 +531,7 @@ def test_ll_deepep_deepgemm_moe(
     assert not is_deep_gemm_e8m0_used()
 
     m, n, k = mnk
-    current_platform.seed_everything(7)
+    set_random_seed(7)
 
     if topk > num_experts:
         pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")

tests/kernels/moe/test_deepep_moe.py

@@ -20,8 +20,8 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularK
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
 )
-from vllm.platforms import current_platform
 from vllm.utils.import_utils import has_deep_ep
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.worker.workspace import init_workspace_manager
 
 from ...utils import multi_gpu_test
@@ -446,7 +446,7 @@ def test_deep_ep_moe(
     low_latency_mode = False
     use_fp8_dispatch = False
 
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     world_size, dp_size = world_dp_size
     config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
 
@@ -507,7 +507,7 @@ def test_low_latency_deep_ep_moe(
             f"hidden sizes {DeepEPLLPrepareAndFinalize.SUPPORTED_HIDDEN_SIZES}"
         )
 
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     world_size, dp_size = world_dp_size
     config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
 

tests/kernels/moe/test_flashinfer.py

@@ -22,6 +22,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
 from vllm.model_executor.layers.quantization.utils.fp8_utils import input_to_float8
 from vllm.model_executor.models.llama4 import Llama4MoE
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 try:
     from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
@@ -158,7 +159,7 @@ def test_flashinfer_per_tensor_moe_fp8_no_graph(
 ):
     if not current_platform.has_device_capability(100):
         pytest.skip("Test is only supported for sm >= 100")
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=True)
@@ -222,7 +223,7 @@ def test_flashinfer_cutlass_moe_fp8_no_graph(
     monkeypatch,
     workspace_init,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
     with set_current_vllm_config(vllm_config):
         td = TestData.make_moe_tensors_8bit(

tests/kernels/moe/test_flashinfer_moe.py

@@ -23,6 +23,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
 from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
 from vllm.platforms import current_platform
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.torch_utils import set_random_seed
 
 if not has_flashinfer_cutlass_fused_moe() or not current_platform.has_device_capability(
     100
@@ -60,7 +61,7 @@ def test_flashinfer_fp4_moe_no_graph(
     activation: str,
     workspace_init,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
     with set_current_vllm_config(
         VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
     ):

tests/kernels/moe/test_grouped_topk.py

@@ -19,6 +19,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
     fused_grouped_topk,
 )
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 
 @pytest.mark.skipif(
@@ -52,7 +53,7 @@ def test_grouped_topk(
     )
     get_cached_compilation_config.cache_clear()
 
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     hidden_states = torch.randn((n_token, n_hidden), dtype=dtype, device="cuda")
     gating_output = torch.randn((n_token, n_expert), dtype=dtype, device="cuda")
     e_score_correction_bias = torch.randn(

tests/kernels/moe/test_modular_kernel_combinations.py

@@ -15,7 +15,7 @@ from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.platforms import current_platform
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
 from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
-from vllm.utils.torch_utils import cuda_device_count_stateless
+from vllm.utils.torch_utils import cuda_device_count_stateless, set_random_seed
 from vllm.v1.worker.workspace import init_workspace_manager
 
 from .modular_kernel_tools.common import (
@@ -82,7 +82,7 @@ def rank_worker(
     device = torch.device(f"cuda:{pgi.local_rank}")
     init_workspace_manager(device)
 
-    current_platform.seed_everything(pgi.rank)
+    set_random_seed(pgi.rank)
 
     # sanity check
     from vllm import envs

tests/kernels/moe/test_modular_oai_triton_moe.py

@@ -34,6 +34,7 @@ from vllm.model_executor.layers.fused_moe.prepare_finalize import (
 )
 from vllm.model_executor.layers.utils import shuffle_weight
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 MNK = [
     (1, 512, 384),
@@ -211,7 +212,7 @@ def test_oai_triton_moe(
     unfused: bool,
     workspace_init,
 ):
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     (
         w1,
         w2,

tests/kernels/moe/test_moe.py

@@ -60,6 +60,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import quantize_w
 from vllm.model_executor.models.mixtral import MixtralMoE
 from vllm.platforms import current_platform
 from vllm.scalar_type import ScalarType, scalar_types
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.worker.workspace import init_workspace_manager
 
 NUM_EXPERTS = [8, 64, 192]
@@ -234,7 +235,7 @@ def test_fused_moe(
     monkeypatch,
     workspace_init,
 ):
-    current_platform.seed_everything(7)
+    set_random_seed(7)
 
     monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(chunk_size))
 

tests/kernels/moe/test_moe_align_block_size.py

@@ -14,12 +14,13 @@ from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
 )
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import round_up
+from vllm.utils.torch_utils import set_random_seed
 
 NUM_TOKENS = [1, 3, 256, 2256, 4096]
 NUM_EXPERTS = [32, 160, 256, 257]
 TOP_KS = [1, 2, 16, 32]
 BLOCK_SIZES = [32, 128]
-current_platform.seed_everything(0)
+set_random_seed(0)
 
 
 def _group_tokens_by_expert(

tests/kernels/moe/test_moe_permute_unpermute.py

@@ -17,11 +17,12 @@ from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
     moe_unpermute,
 )
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 NUM_EXPERTS = [16, 64, 256]
 TOP_KS = [2, 6, 8]
 EP_SIZE = [1, 4, 16]
-current_platform.seed_everything(0)
+set_random_seed(0)
 
 if current_platform.is_rocm():
     pytest.skip(
@@ -226,7 +227,7 @@ def test_moe_permute_unpermute(
         n_local_expert, expert_map, _ = determine_expert_map(ep_size, ep_rank, n_expert)
         expert_map = expert_map.cuda()
     start_expert = n_local_expert * ep_rank
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     hidden_states = torch.randn((n_token, n_hidden), device="cuda").to(dtype)
     gating_output = torch.randn((n_token, n_expert), device="cuda").to(dtype)
     topk_weights, topk_ids, token_expert_indices = fused_topk(