[Platform] Deprecate seed_everything (#31659)

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

tests/kernels/test_fla_layernorm_guard.py

@@ -10,7 +10,7 @@ from vllm.model_executor.layers.fla.ops.layernorm_guard import (
     layernorm_fn,
     rms_norm_ref,
 )
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 
 def layer_norm_ref(
@@ -114,7 +114,7 @@ def test_layer_norm_fwd_basic(
     is_rms_norm: bool,
 ) -> None:
     """Test basic layer norm forward pass without z (gate) tensor."""
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
     device = torch.device("cuda:0")
 
     # Create inputs
@@ -156,7 +156,7 @@ def test_layer_norm_fwd_with_gate(
     is_rms_norm: bool,
 ) -> None:
     """Test layer norm forward pass with z (gate) tensor."""
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
 
     # Create inputs
@@ -213,7 +213,7 @@ def test_layer_norm_fwd_with_groups(
             f"hidden_size {hidden_size} not divisible by group_size {group_size}"
         )
 
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
 
     # Create inputs
@@ -253,7 +253,7 @@ def test_layer_norm_rows_per_block(
     dtype: torch.dtype,
 ) -> None:
     """Test that rows_per_block logic works correctly for various M sizes."""
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
     hidden_size = 1024
 
@@ -278,7 +278,7 @@ def test_layer_norm_rows_per_block(
 def test_strided_input(dtype: torch.dtype) -> None:
     """Test that the kernel handles non-contiguous (strided)
     inputs correctly."""
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
     num_tokens = 128
     hidden_size = 1024
@@ -318,7 +318,7 @@ def test_output_buffer_provided(
     dtype: torch.dtype,
 ) -> None:
     """Test that the kernel works when an output buffer is provided."""
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
 
     # Create inputs
@@ -359,7 +359,7 @@ def test_multidimensional_input(
     dtype: torch.dtype,
 ) -> None:
     """Test that the autograd function handles multidimensional inputs."""
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     device = torch.device("cuda:0")
     hidden_size = shape[-1]
 

tests/lora/test_fused_moe_lora_kernel.py

@@ -18,8 +18,8 @@ from vllm.distributed.parallel_state import (
     get_tensor_model_parallel_world_size,
 )
 from vllm.lora.ops.triton_ops import fused_moe_lora
-from vllm.platforms import current_platform
 from vllm.utils.network_utils import get_open_port
+from vllm.utils.torch_utils import set_random_seed
 
 
 @pytest.fixture(autouse=True)
@@ -265,7 +265,7 @@ def test_fused_moe_lora_kernel(
     seed,
 ):
     torch.set_default_device(device)
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
     # the number of randomly generated sentences.
     num_sequences = 10
     # generate data
@@ -358,7 +358,7 @@ def test_fused_moe_lora_kernel_fully_sharded(
     seed,
     column_parallel,
 ):
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
     # the number of randomly generated sentences.
     num_sequences = 10
     # generate data
@@ -415,7 +415,7 @@ def use_fused_moe_lora_kernel_tensor_parallel(
     def _get_shard_slice(shard_size):
         return slice(local_rank * shard_size, (local_rank + 1) * shard_size)
 
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     device = torch.device(f"cuda:{local_rank}")
     torch.cuda.set_device(device)

tests/lora/test_layers.py

@@ -43,8 +43,8 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding,
     get_masked_input_and_mask,
 )
-from vllm.model_executor.utils import set_random_seed
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 from .utils import DummyLoRAManager
 

tests/lora/test_punica_ops.py

@@ -9,7 +9,7 @@ import vllm.lora.ops.torch_ops as torch_ops
 import vllm.lora.ops.triton_ops as triton_ops
 from vllm.lora.ops.triton_ops import LoRAKernelMeta
 from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
-from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 from .utils import PunicaTensors, assert_close, generate_data_for_nslices
 
@@ -395,7 +395,7 @@ def test_kernels(
     Tests LoRA kernels.
     """
     torch.set_default_device(device)
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     if op_type == "shrink":
         check_lora_shrink_kernel(
@@ -447,7 +447,7 @@ def test_kernels_hidden_size(
     Tests SGMV and LoRA kernels.
     """
     torch.set_default_device(device)
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     if op_type == "shrink":
         check_lora_shrink_kernel(

tests/models/test_vision.py

@@ -21,6 +21,7 @@ from vllm.model_executor.models.vision import (
 from vllm.platforms import current_platform
 from vllm.utils.network_utils import get_open_port
 from vllm.utils.system_utils import update_environment_variables
+from vllm.utils.torch_utils import set_random_seed
 
 pytestmark = pytest.mark.cpu_test
 
@@ -98,7 +99,7 @@ def run_dp_sharded_vision_model_vs_direct(
     """
 
     # Set random seed for reproducibility
-    current_platform.seed_everything(0)
+    set_random_seed(0)
 
     device = f"{current_platform.device_name}:{local_rank}"
     current_platform.set_device(device)
@@ -284,7 +285,7 @@ def run_dp_sharded_mrope_vision_model_vs_direct(
     calling the model directly.
     """
     # Set random seed for reproducibility
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     device = f"{current_platform.device_name}:{local_rank}"
     current_platform.set_device(device)
     torch.set_default_device(device)
@@ -408,7 +409,7 @@ def run_dp_sharded_mrope_vision_model_uneven_load_worker(
 ):
     """Test run_dp_sharded_mrope_vision_model with uneven load distribution."""
     # Set up distributed environment
-    current_platform.seed_everything(123)
+    set_random_seed(123)
     device = f"{current_platform.device_name}:{local_rank}"
     current_platform.set_device(device)
     torch.set_default_device(device)

tests/v1/attention/test_attention_backends.py

@@ -19,7 +19,11 @@ from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.config import ModelConfig
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import cdiv
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, is_torch_equal_or_newer
+from vllm.utils.torch_utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    is_torch_equal_or_newer,
+    set_random_seed,
+)
 from vllm.v1.attention.backends.utils import (
     CommonAttentionMetadata,
     set_kv_cache_layout,
@@ -320,7 +324,7 @@ def _test_backend_correctness(
     multiple GPUs. This tests that backends work correctly with different
     head counts.
     """
-    current_platform.seed_everything(42)
+    set_random_seed(42)
 
     hf_config_override = None
     if tensor_parallel_size > 1:

tests/v1/kv_offload/test_cpu_gpu.py

@@ -7,6 +7,7 @@ import pytest
 import torch
 
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
 from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
 from vllm.v1.kv_offload.worker.cpu_gpu import CpuGpuOffloadingHandlers
@@ -62,7 +63,7 @@ def test_transfer(
     seed: int,
     device: str,
 ) -> None:
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     # create per-layer GPU KV caches based on available attn_backends
     attn_backends_list = BACKENDS_TO_TEST

tests/v1/tpu/test_mha_attn.py

@@ -15,6 +15,7 @@ import torch_xla.core.xla_model
 from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
 from vllm.attention.selector import _cached_get_attn_backend
 from vllm.platforms import current_platform
+from vllm.utils.torch_utils import set_random_seed
 
 
 @pytest.fixture(autouse=True)
@@ -63,7 +64,7 @@ def test_mha_attn_forward(
     head_size: int,
     device: str,
 ):
-    current_platform.seed_everything(0)
+    set_random_seed(0)
     # These are expected to be f32
     q = torch.randn(batch_size, seq_len, num_heads * head_size, device=device)
     k = torch.randn(batch_size, seq_len, num_kv_heads * head_size, device=device)

tests/v1/worker/test_gpu_model_runner.py

@@ -26,6 +26,7 @@ from vllm.platforms import current_platform
 from vllm.sampling_params import SamplingParams
 from vllm.utils.mem_constants import GiB_bytes
 from vllm.utils.system_utils import update_environment_variables
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.core.kv_cache_utils import estimate_max_model_len, get_kv_cache_configs
 from vllm.v1.core.sched.output import CachedRequestData, NewRequestData, SchedulerOutput
 from vllm.v1.kv_cache_interface import (
@@ -776,7 +777,7 @@ def test_hybrid_attention_mamba_tensor_shapes():
     will not corrupt an attention block and vice versa
     """
 
-    current_platform.seed_everything(42)
+    set_random_seed(42)
 
     update_environment_variables(
         {

vllm/model_executor/__init__.py

@@ -2,10 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from vllm.model_executor.parameter import BasevLLMParameter, PackedvLLMParameter
-from vllm.model_executor.utils import set_random_seed
 
 __all__ = [
-    "set_random_seed",
     "BasevLLMParameter",
     "PackedvLLMParameter",
 ]

vllm/model_executor/utils.py

@@ -10,12 +10,6 @@ import torch
 from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 
-def set_random_seed(seed: int | None) -> None:
-    from vllm.platforms import current_platform
-
-    current_platform.seed_everything(seed)
-
-
 def set_weight_attrs(
     weight: torch.Tensor,
     weight_attrs: dict[str, Any] | None,

vllm/platforms/interface.py

@@ -372,6 +372,10 @@ class Platform:
 
         Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
         """
+        logger.info_once(
+            "`seed_everything` is deprecated. It will be removed in v0.14.0 or later. "
+            "Please use `vllm.utils.torch_utils.set_random_seed` instead."
+        )
         if seed is not None:
             random.seed(seed)
             np.random.seed(seed)

vllm/utils/torch_utils.py

@@ -3,6 +3,7 @@
 import contextlib
 import importlib.metadata
 import os
+import random
 import threading
 from collections.abc import Callable, Collection
 from functools import lru_cache
@@ -278,6 +279,13 @@ def kv_cache_dtype_str_to_dtype(
     return STR_DTYPE_TO_TORCH_DTYPE[kv_cache_dtype]
 
 
+def set_random_seed(seed: int | None) -> None:
+    if seed is not None:
+        random.seed(seed)
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+
+
 def create_kv_caches_with_random_flash(
     num_blocks: int,
     block_size: int,
@@ -290,9 +298,7 @@ def create_kv_caches_with_random_flash(
     device: str | None = "cuda",
     cache_layout: str | None = "NHD",
 ) -> tuple[list[torch.Tensor], list[torch.Tensor]]:
-    from vllm.platforms import current_platform
-
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
     generic_kv_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
@@ -335,9 +341,8 @@ def create_kv_caches_with_random(
         raise ValueError(
             f"Does not support key cache of type fp8 with head_size {head_size}"
         )
-    from vllm.platforms import current_platform
 
-    current_platform.seed_everything(seed)
+    set_random_seed(seed)
 
     dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
 

vllm/v1/worker/cpu_worker.py

@@ -10,10 +10,10 @@ import torch
 from vllm import envs
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
-from vllm.model_executor.utils import set_random_seed
 from vllm.platforms import CpuArchEnum, current_platform
 from vllm.platforms.cpu import CpuPlatform, LogicalCPUInfo
 from vllm.profiler.wrapper import TorchProfilerWrapper
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.worker.cpu_model_runner import CPUModelRunner
 from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
 

vllm/v1/worker/gpu_worker.py

@@ -34,7 +34,6 @@ from vllm.distributed.parallel_state import (
 )
 from vllm.logger import init_logger
 from vllm.lora.request import LoRARequest
-from vllm.model_executor import set_random_seed
 from vllm.model_executor.models.interfaces import is_mixture_of_experts
 from vllm.model_executor.warmup.kernel_warmup import kernel_warmup
 from vllm.platforms import current_platform
@@ -43,6 +42,7 @@ from vllm.sequence import IntermediateTensors
 from vllm.tasks import SupportedTask
 from vllm.utils.mem_constants import GiB_bytes
 from vllm.utils.mem_utils import MemorySnapshot, memory_profiling
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
 from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec

vllm/v1/worker/tpu_worker.py

@@ -20,12 +20,11 @@ from vllm.distributed.kv_transfer import (
 )
 from vllm.logger import init_logger
 from vllm.lora.request import LoRARequest
-from vllm.model_executor import set_random_seed
 from vllm.platforms import current_platform
 from vllm.platforms.tpu import USE_TPU_INFERENCE
 from vllm.tasks import SupportedTask
 from vllm.utils.math_utils import cdiv
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
+from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
 from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
 from vllm.v1.kv_cache_interface import AttentionSpec, KVCacheConfig, KVCacheSpec
 from vllm.v1.outputs import ModelRunnerOutput

vllm/v1/worker/xpu_worker.py

@@ -9,9 +9,9 @@ import torch.distributed
 from vllm.config import VllmConfig
 from vllm.distributed import get_world_group
 from vllm.logger import init_logger
-from vllm.model_executor import set_random_seed
 from vllm.platforms import current_platform
 from vllm.profiler.wrapper import TorchProfilerWrapper
+from vllm.utils.torch_utils import set_random_seed
 from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
 from vllm.v1.worker.xpu_model_runner import XPUModelRunner