[CI/Build] Avoid CUDA initialization (#8534)

benchmarks/kernels/benchmark_layernorm.py

-import random
 import time
 
 import torch
 
 from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        seed_everything)
 
 
 @torch.inference_mode()
@@ -16,10 +16,7 @@ def main(num_tokens: int,
          do_profile: bool = False,
          num_warmup_iters: int = 5,
          num_iters: int = 100) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device("cuda")
 
     layer = RMSNorm(hidden_size).to(dtype=dtype)

benchmarks/kernels/benchmark_moe.py

@@ -10,7 +10,7 @@ from ray.experimental.tqdm_ray import tqdm
 from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, seed_everything
 
 
 class BenchmarkConfig(TypedDict):
@@ -166,7 +166,7 @@ class BenchmarkWorker:
 
     def __init__(self, seed: int) -> None:
         torch.set_default_device("cuda")
-        torch.cuda.manual_seed_all(seed)
+        seed_everything(seed)
         self.seed = seed
 
     def benchmark(
@@ -180,7 +180,7 @@ class BenchmarkWorker:
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
     ) -> Tuple[Dict[str, int], float]:
-        torch.cuda.manual_seed_all(self.seed)
+        seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
                                          use_int8_w8a16=use_int8_w8a16,
                                          use_fp8_w8a8=use_fp8_w8a8)

benchmarks/kernels/benchmark_paged_attention.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        create_kv_caches_with_random)
+                        create_kv_caches_with_random, seed_everything)
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -28,10 +28,7 @@ def main(
     device: str = "cuda",
     kv_cache_dtype: Optional[str] = None,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     scale = float(1.0 / (head_size**0.5))
     query = torch.empty(num_seqs,

benchmarks/kernels/benchmark_quant.py

-import random
 import time
 
 import torch
 
 from vllm import _custom_ops as ops
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
+                        seed_everything)
 
 
 @torch.inference_mode()
@@ -17,10 +17,7 @@ def main(num_tokens: int,
          do_profile: bool = False,
          num_warmup_iters: int = 5,
          num_iters: int = 100) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device("cuda")
 
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)

benchmarks/kernels/benchmark_rope.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
                                                          get_rope)
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, seed_everything
 
 
 def benchmark_rope_kernels_multi_lora(
@@ -22,9 +22,7 @@ def benchmark_rope_kernels_multi_lora(
     max_position: int = 8192,
     base: int = 10000,
 ) -> None:
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     if rotary_dim is None:
         rotary_dim = head_size

tests/kernels/test_activation.py

@@ -7,6 +7,7 @@ from tests.kernels.utils import opcheck
 from vllm.model_executor.layers.activation import (FastGELU, GeluAndMul,
                                                    NewGELU, QuickGELU,
                                                    SiluAndMul)
+from vllm.utils import seed_everything
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -34,9 +35,7 @@ def test_act_and_mul(
     seed: int,
     device: str,
 ) -> None:
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     x = torch.randn(num_tokens, 2 * d, dtype=dtype)
     if activation == "silu":
@@ -77,9 +76,7 @@ def test_activation(
     seed: int,
     device: str,
 ) -> None:
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     x = torch.randn(num_tokens, d, dtype=dtype)
     layer = activation[0]()

tests/kernels/test_attention.py

@@ -6,7 +6,7 @@ import torch
 
 from tests.kernels.utils import opcheck
 from vllm import _custom_ops as ops
-from vllm.utils import get_max_shared_memory_bytes, is_hip
+from vllm.utils import get_max_shared_memory_bytes, is_hip, seed_everything
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -139,10 +139,8 @@ def test_paged_attention(
 ) -> None:
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+
+    seed_everything(seed)
     torch.set_default_device(device)
     scale = float(1.0 / (head_size**0.5))
     num_query_heads, num_kv_heads = num_heads
@@ -354,10 +352,7 @@ def test_paged_attention_rocm(
     seed: int,
     device: str,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     scale = float(1.0 / (head_size**0.5))
     num_query_heads, num_kv_heads = num_heads
@@ -506,10 +501,7 @@ def test_multi_query_kv_attention(
     seed: int,
     device: str,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
     # As the xformers library is already tested with its own tests, we can use

tests/kernels/test_attention_selector.py

@@ -45,7 +45,7 @@ def test_flash_attn(monkeypatch):
     override_backend_env_variable(monkeypatch, STR_FLASH_ATTN_VAL)
 
     # Unsupported CUDA arch
-    with patch("torch.cuda.get_device_capability", return_value=[7, 5]):
+    with patch("torch.cuda.get_device_capability", return_value=(7, 5)):
         backend = which_attn_to_use(8, 16, 8, None, torch.float16, None, 16)
         assert backend.name != STR_FLASH_ATTN_VAL
 

tests/kernels/test_awq_triton.py

@@ -7,6 +7,7 @@ import torch
 
 from vllm.model_executor.layers.quantization.awq_triton import (
     AWQ_TRITON_SUPPORTED_GROUP_SIZES, awq_dequantize_triton, awq_gemm_triton)
+from vllm.utils import seed_everything
 
 device = "cuda"
 
@@ -79,7 +80,7 @@ def test_dequantize(qweight_rows, qweight_cols, group_size):
     zeros_cols = qweight_cols
     zeros_dtype = torch.int32
 
-    torch.manual_seed(0)
+    seed_everything(0)
 
     qweight = torch.randint(0,
                             torch.iinfo(torch.int32).max,
@@ -133,7 +134,7 @@ def test_gemm(N, K, M, splitK, group_size):
     qzeros_rows = scales_rows
     qzeros_cols = qweight_cols
 
-    torch.manual_seed(0)
+    seed_everything(0)
 
     input = torch.rand((input_rows, input_cols),
                        dtype=input_dtype,

tests/kernels/test_blocksparse_attention.py

@@ -7,7 +7,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.attention.ops.blocksparse_attention.interface import (
     LocalStridedBlockSparseAttn)
-from vllm.utils import get_max_shared_memory_bytes, is_hip
+from vllm.utils import get_max_shared_memory_bytes, is_hip, seed_everything
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -172,10 +172,7 @@ def test_paged_attention(
     blocksparse_block_size: int,
     blocksparse_head_sliding_step: int,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     scale = float(1.0 / (head_size**0.5))
     num_query_heads, num_kv_heads = num_heads
@@ -386,10 +383,7 @@ def test_varlen_blocksparse_attention_prefill(
     seed: int,
     device: str,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
     # As the xformers library is already tested with its own tests, we can use

tests/kernels/test_cache.py

@@ -6,6 +6,7 @@ import torch
 
 from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
 from vllm import _custom_ops as ops
+from vllm.utils import seed_everything
 
 COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]
 DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -55,10 +56,7 @@ def test_copy_blocks(
 ) -> None:
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     # Generate random block mappings where each source block is mapped to two
     # destination blocks.
@@ -134,10 +132,7 @@ def test_reshape_and_cache(
 ) -> None:
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     # Create a random slot mapping.
     num_slots = block_size * num_blocks
@@ -229,9 +224,7 @@ def test_reshape_and_cache_flash(
     device: str,
     kv_cache_dtype: str,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
 
     # Create a random slot mapping.
@@ -345,10 +338,8 @@ def test_swap_blocks(
         pytest.skip()
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+
+    seed_everything(seed)
 
     src_device = device if direction[0] == "cuda" else 'cpu'
     dst_device = device if direction[1] == "cuda" else 'cpu'
@@ -417,9 +408,7 @@ def test_fp8_e4m3_conversion(
     seed: int,
     device: str,
 ) -> None:
-    random.seed(seed)
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     low = -224.0
     high = 224.0

tests/kernels/test_causal_conv1d.py

@@ -7,6 +7,7 @@ from einops import rearrange
 
 from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
     causal_conv1d_fn, causal_conv1d_update)
+from vllm.utils import seed_everything
 
 
 def causal_conv1d_ref(
@@ -104,7 +105,7 @@ def test_causal_conv1d(batch, dim, seqlen, width, has_bias, silu_activation,
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    torch.random.manual_seed(0)
+    seed_everything(0)
     if not channel_last:
         x = torch.randn(batch,
                         4096 + dim + 64,
@@ -175,7 +176,7 @@ def test_causal_conv1d_update(batch, dim, width, has_bias, silu_activation,
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    torch.random.manual_seed(0)
+    seed_everything(0)
     batch = 2
     x = torch.randn(batch, dim, device=device, dtype=itype)
     conv_state = torch.randn(batch, dim, width, device=device, dtype=itype)

tests/kernels/test_cutlass.py

@@ -15,9 +15,6 @@ CUDA_DEVICES = [
     f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
 ]
 
-capability = current_platform.get_device_capability()
-capability = capability[0] * 10 + capability[1]
-
 
 def to_fp8(tensor: torch.Tensor):
     finfo = torch.finfo(torch.float8_e4m3fn)
@@ -119,7 +116,7 @@ def cutlass_int8_gemm_helper(m: int,
 @pytest.mark.parametrize("per_act_token", [True, False])
 @pytest.mark.parametrize("per_out_ch", [True, False])
 @pytest.mark.parametrize("use_bias", [True, False])
-@pytest.mark.skipif(capability < 89,
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
                     reason="FP8 is not supported on this GPU type.")
 def test_cutlass_fp8_gemm(m: int, n: int, k: int, per_act_token: bool,
                           per_out_ch: bool, use_bias: bool):
@@ -157,7 +154,7 @@ def test_cutlass_int8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
 @pytest.mark.parametrize("per_out_ch", [True, False])
 @pytest.mark.parametrize("out_dtype", [torch.bfloat16, torch.float16])
 @pytest.mark.parametrize("use_bias", [True, False])
-@pytest.mark.skipif(capability < 89,
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
                     reason="FP8 is not supported on this GPU type.")
 def test_cutlass_fp8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
                                        out_dtype: Type[torch.dtype],
@@ -175,7 +172,7 @@ def test_cutlass_fp8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
 @pytest.mark.parametrize("per_out_ch", [True, False])
 @pytest.mark.parametrize("use_bias", [True, False])
 @pytest.mark.parametrize("device", CUDA_DEVICES)
-@pytest.mark.skipif(capability < 89,
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
                     reason="FP8 is not supported on this GPU type.")
 def test_cutlass_fp8_gemm_devices(per_act_token: bool, per_out_ch: bool,
                                   use_bias: bool, device: str):
@@ -207,7 +204,7 @@ def test_cutlass_int8_gemm_devices(per_act_token: bool, per_out_ch: bool,
 @pytest.mark.parametrize("per_act_token", [True, False])
 @pytest.mark.parametrize("per_out_ch", [True, False])
 @pytest.mark.parametrize("use_bias", [True, False])
-@pytest.mark.skipif(capability < 89,
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
                     reason="FP8 is not supported on this GPU type.")
 def test_cutlass_fp8_gemm_m_sweep(per_act_token: bool, per_out_ch: bool,
                                   use_bias: bool):

tests/kernels/test_flash_attn.py

@@ -4,6 +4,7 @@ import pytest
 import torch
 
 import vllm.attention.backends.flash_attn  # noqa: F401
+from vllm.utils import seed_everything
 
 NUM_HEADS = [(4, 4), (8, 2), (16, 2)]
 HEAD_SIZES = [128, 256]
@@ -87,7 +88,7 @@ def test_flash_attn_with_paged_kv(
     num_blocks: int,
 ) -> None:
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]
@@ -174,7 +175,7 @@ def test_varlen_with_paged_kv(
     num_blocks: int,
 ) -> None:
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]

tests/kernels/test_flashinfer.py

@@ -4,6 +4,8 @@ import flashinfer
 import pytest
 import torch
 
+from vllm.utils import seed_everything
+
 NUM_HEADS = [(16, 16), (32, 8), (64, 8), (6, 1)]
 HEAD_SIZES = [128, 256]
 BLOCK_SIZES = [16, 32]
@@ -82,7 +84,7 @@ def test_flashinfer_decode_with_paged_kv(
     soft_cap: Optional[float],
 ) -> None:
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]
@@ -168,7 +170,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
                                           block_size: int,
                                           soft_cap: Optional[float]) -> None:
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]
@@ -266,7 +268,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
         head_size: int, dtype: torch.dtype, block_size: int,
         soft_cap: Optional[float]) -> None:
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]
@@ -379,7 +381,7 @@ def test_flashinfer_decode_with_paged_fp8_kv(
 ) -> None:
     # test doesn't work for num_heads = (16,16)
     torch.set_default_device("cuda")
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]

tests/kernels/test_fp8_quant.py

@@ -5,6 +5,7 @@ import vllm._custom_ops as ops
 from tests.kernels.quant_utils import (FP8_DTYPE,
                                        ref_dynamic_per_tensor_fp8_quant,
                                        ref_dynamic_per_token_quant)
+from vllm.utils import seed_everything
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
 HIDDEN_SIZES = [1, 2, 3, 4, 16, 67, 768, 2048, 5120, 5137, 8192,
@@ -24,8 +25,7 @@ SEEDS = [0]
 def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
                                      dtype: torch.dtype, scale_ub: bool,
                                      seed: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype,
                    device="cuda") + 1e-6  # avoid nans
@@ -49,8 +49,7 @@ def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
 @torch.inference_mode()
 def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
                                       dtype: torch.dtype, seed: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
 
@@ -67,8 +66,7 @@ def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
 @torch.inference_mode()
 @pytest.mark.parametrize("seed", SEEDS)
 def test_fp8_quant_large(seed: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     num_tokens = 1024000  # Mistral-Nemo's max_position_embeddings
     hidden_size = 1152  # Smallest hidden_size to reproduce the error

tests/kernels/test_gguf.py

@@ -7,6 +7,7 @@ from gguf import GGMLQuantizationType, GGUFReader, ReaderTensor, dequantize
 from huggingface_hub import snapshot_download
 
 import vllm._custom_ops as ops
+from vllm.utils import seed_everything
 
 GGUF_SAMPLE = snapshot_download("Isotr0py/test-gguf-sample")
 
@@ -74,7 +75,7 @@ def test_dequantize(hidden_size: int, dtype: torch.dtype,
 @torch.inference_mode()
 def test_mmvq(hidden_size: int, dtype: torch.dtype,
               quant_type: GGMLQuantizationType):
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
 
     tensors = get_gguf_sample_tensors(hidden_size, quant_type)
     x = torch.rand((1, hidden_size), dtype=dtype, device="cuda")
@@ -110,7 +111,7 @@ def test_mmvq(hidden_size: int, dtype: torch.dtype,
 @torch.inference_mode()
 def test_mmq(num_tokens: int, hidden_size: int, dtype: torch.dtype,
              quant_type: GGMLQuantizationType):
-    torch.cuda.manual_seed_all(0)
+    seed_everything(0)
 
     tensors = get_gguf_sample_tensors(hidden_size, quant_type)
     x = torch.rand((num_tokens, hidden_size), dtype=dtype, device="cuda")

tests/kernels/test_int8_quant.py

@@ -4,6 +4,7 @@ import torch
 from tests.kernels.quant_utils import ref_dynamic_per_token_quant
 from tests.kernels.utils import opcheck
 from vllm._custom_ops import scaled_int8_quant
+from vllm.utils import seed_everything
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
 HIDDEN_SIZES = [16, 67, 768, 2048, 5120, 5137, 8192,
@@ -44,8 +45,7 @@ def opcheck_int8_quant_dynamic(output, input, symmetric=True):
 @torch.inference_mode()
 def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
                                    dtype: torch.dtype, seed: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
 
@@ -68,8 +68,7 @@ def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
 @torch.inference_mode()
 def test_dynamic_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
                                        dtype: torch.dtype, seed: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     int8_traits = torch.iinfo(torch.int8)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype,
@@ -113,8 +112,7 @@ def test_dynamic_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
 def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
                                   dtype: torch.dtype, seed: int,
                                   scale: float) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     int8_traits = torch.iinfo(torch.int8)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
@@ -140,8 +138,7 @@ def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
 def test_static_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
                                       dtype: torch.dtype, seed: int,
                                       scale: float, azp: int) -> None:
-    torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     int8_traits = torch.iinfo(torch.int8)
 
     x = torch.rand(num_tokens, hidden_size, dtype=dtype,

tests/kernels/test_layernorm.py

@@ -3,6 +3,7 @@ import torch
 
 from tests.kernels.utils import opcheck
 from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.utils import seed_everything
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
 NUM_TOKENS = [7, 83, 4096]  # Arbitrary values for testing
@@ -30,9 +31,7 @@ def test_rms_norm(
     seed: int,
     device: str,
 ) -> None:
-    torch.random.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed(seed)
+    seed_everything(seed)
     torch.set_default_device(device)
     layer = RMSNorm(hidden_size).to(dtype=dtype)
     layer.weight.data.normal_(mean=1.0, std=0.1)

tests/kernels/test_machete_gemm.py

@@ -48,7 +48,7 @@ WTYPE_ZEROPOINTS = [
 #  `is_quant_method_supported` conflates kernels with quantization methods
 #  an assumption which is breaking down as quantizations methods can have
 #  have kernels and some kernels support multiple quantization methods.
-IS_SUPPORTED_BY_GPU = current_platform.get_device_capability()[0] >= 9
+IS_SUPPORTED_BY_GPU = current_platform.has_device_capability(90)
 
 
 def rand_data(shape, dtype=torch.float16):