[Platform] Deprecate seed_everything (#31659)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
@@ -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:
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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)
|
||||
|
||||
@@ -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(
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user