[Platform] Deprecate seed_everything (#31659)

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

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