From ec68d53b2b75eb5480270c67676b126079998f5a Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Fri, 10 Apr 2026 13:43:50 +0800 Subject: [PATCH] Add platform manual_seed_all API (#38468) Signed-off-by: Yan Ma --- benchmarks/kernels/benchmark_moe_align_block_size.py | 3 ++- .../kernels/deepgemm/benchmark_fp8_block_dense_gemm.py | 5 ++--- tests/kernels/attention/test_lightning_attn.py | 6 ------ tests/kernels/core/test_fused_quant_layernorm.py | 5 ++--- tests/kernels/moe/test_marlin_vs_trtllm_mxint4.py | 5 +++-- tests/kernels/moe/test_moe.py | 10 +++++----- .../moe/test_shared_fused_moe_routed_transform.py | 5 ++--- tests/kernels/test_fused_quant_activation.py | 5 ++--- .../language/pooling/test_token_classification.py | 9 ++------- tests/v1/e2e/spec_decode/test_lora_with_spec_decode.py | 9 ++------- tests/v1/spec_decode/test_tree_attention.py | 4 ++-- tools/pre_commit/check_torch_cuda.py | 9 +++++++++ vllm/platforms/cpu.py | 4 ++++ vllm/platforms/cuda.py | 4 ++++ vllm/platforms/interface.py | 5 +++++ vllm/platforms/rocm.py | 4 ++++ vllm/platforms/xpu.py | 4 ++++ vllm/utils/torch_utils.py | 5 +++-- 18 files changed, 57 insertions(+), 44 deletions(-) diff --git a/benchmarks/kernels/benchmark_moe_align_block_size.py b/benchmarks/kernels/benchmark_moe_align_block_size.py index 5f9a131f79..a340500379 100644 --- a/benchmarks/kernels/benchmark_moe_align_block_size.py +++ b/benchmarks/kernels/benchmark_moe_align_block_size.py @@ -9,6 +9,7 @@ from vllm.model_executor.layers.fused_moe.moe_align_block_size import ( moe_align_block_size, ) from vllm.triton_utils import triton +from vllm.utils.torch_utils import set_random_seed def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor: @@ -44,7 +45,7 @@ configs = list( def benchmark(num_tokens, num_experts, topk, ep_size, provider): """Benchmark function for Triton.""" block_size = 256 - torch.cuda.manual_seed_all(0) + set_random_seed(0) topk_ids = get_topk_ids(num_tokens, num_experts, topk) e_map = None diff --git a/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py b/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py index 4384d3e568..c9aaef284d 100644 --- a/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py +++ b/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py @@ -16,6 +16,7 @@ from vllm.utils.deep_gemm import ( fp8_gemm_nt, per_block_cast_to_fp8, ) +from vllm.utils.torch_utils import set_random_seed def benchmark_shape( @@ -235,9 +236,7 @@ def run_benchmarks(verbose: bool = False): torch.backends.cudnn.allow_tf32 = True # Set seeds for reproducibility - torch.manual_seed(42) - torch.cuda.manual_seed(42) - + set_random_seed(42) # Define benchmark shapes (m, n, k) shapes = [ (8, 4096, 7168), diff --git a/tests/kernels/attention/test_lightning_attn.py b/tests/kernels/attention/test_lightning_attn.py index 37fd85ccec..46757cc10b 100644 --- a/tests/kernels/attention/test_lightning_attn.py +++ b/tests/kernels/attention/test_lightning_attn.py @@ -122,8 +122,6 @@ def test_linear_decode_forward_triton( dtype: torch.dtype, ): torch.set_default_device("cuda") - torch.manual_seed(42) - torch.cuda.manual_seed_all(42) set_random_seed(42) base = 0.01 q = base * torch.randn(batch_size, num_heads, 1, head_size, dtype=dtype) @@ -165,8 +163,6 @@ def test_linear_decode_forward_triton_with_padding( dtype: torch.dtype, ): torch.set_default_device("cuda") - torch.manual_seed(42) - torch.cuda.manual_seed_all(42) set_random_seed(42) batch_size = 4 @@ -229,8 +225,6 @@ def test_lightning_attention_reference( dtype: torch.dtype, ): torch.set_default_device("cuda") - torch.manual_seed(42) - torch.cuda.manual_seed_all(42) set_random_seed(42) base = 0.01 diff --git a/tests/kernels/core/test_fused_quant_layernorm.py b/tests/kernels/core/test_fused_quant_layernorm.py index f9c01f4f1e..07d15e3b1d 100644 --- a/tests/kernels/core/test_fused_quant_layernorm.py +++ b/tests/kernels/core/test_fused_quant_layernorm.py @@ -17,6 +17,7 @@ from vllm.model_executor.layers.quantization.utils.int8_utils import ( per_token_group_quant_int8, ) from vllm.platforms import current_platform +from vllm.utils.torch_utils import set_random_seed DTYPES = [torch.bfloat16, torch.float] QUANT_DTYPES = [torch.int8, current_platform.fp8_dtype()] @@ -180,9 +181,7 @@ def test_rms_norm( device: str, strided_input: bool, ) -> None: - torch.random.manual_seed(seed) - if torch.cuda.is_available(): - torch.cuda.manual_seed(seed) + set_random_seed(seed) torch.set_default_device(device) torch.accelerator.set_device_index(device) diff --git a/tests/kernels/moe/test_marlin_vs_trtllm_mxint4.py b/tests/kernels/moe/test_marlin_vs_trtllm_mxint4.py index 0ce3d165d0..0ad2cd06ee 100644 --- a/tests/kernels/moe/test_marlin_vs_trtllm_mxint4.py +++ b/tests/kernels/moe/test_marlin_vs_trtllm_mxint4.py @@ -16,6 +16,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_mxint4_moe import ) from vllm.platforms import current_platform from vllm.scalar_type import scalar_types +from vllm.utils.torch_utils import set_random_seed def mxint4_quantize( @@ -134,7 +135,7 @@ def test_marlin_vs_trtllm_mxint4_moe_kimik2(monkeypatch, m, n, k, e, topk, group pytest.importorskip("flashinfer") monkeypatch.setenv("VLLM_USE_FLASHINFER_MOE_INT4", "1") - torch.cuda.manual_seed(0) + set_random_seed(0) dtype = torch.bfloat16 @@ -289,7 +290,7 @@ def test_flashinfer_trtllm_mxint4_moe_wrapper(m, n, k, e, topk): flashinfer_trtllm_mxint4_moe, ) - torch.cuda.manual_seed(0) + set_random_seed(0) dtype = torch.bfloat16 a = torch.randn((m, k), device="cuda", dtype=dtype) * 0.5 diff --git a/tests/kernels/moe/test_moe.py b/tests/kernels/moe/test_moe.py index 5ec2b106e1..def545977e 100644 --- a/tests/kernels/moe/test_moe.py +++ b/tests/kernels/moe/test_moe.py @@ -1031,7 +1031,7 @@ def test_fused_marlin_moe( act_order: bool, is_k_full: bool, ): - torch.cuda.manual_seed(1) + set_random_seed(1) group_size = group_blocks if group_blocks <= 0 else group_blocks * 16 if c_type == scalar_types.float16: @@ -1131,7 +1131,7 @@ def test_fused_marlin_moe( @pytest.mark.skipif(current_platform.is_rocm(), reason="Skip for rocm") @pytest.mark.parametrize("m", [1, 256]) def test_fused_marlin_moe_with_bias(m): - torch.cuda.manual_seed(0) + set_random_seed(0) e, topk = 32, 4 n, k = 2048, 2048 @@ -1213,7 +1213,7 @@ def test_fused_marlin_moe_non_gated( Non-gated activations like relu2 don't have the gate-up projection pattern, so w1 has shape (e, n, k) instead of (e, 2*n, k). """ - torch.cuda.manual_seed(42) + set_random_seed(42) group_size = 16 # NVFP4 group size is_k_full = True @@ -1397,7 +1397,7 @@ def test_cpu_fused_moe_basic( from vllm.model_executor.layers.fused_moe.cpu_fused_moe import CPUFusedMOE device = "cpu" - torch.manual_seed(7) + set_random_seed(7) a = torch.randn((m, k), device=device, dtype=dtype) / 10 w13 = torch.randn((e, 2 * n, k), device=device, dtype=dtype) / 10 @@ -1469,7 +1469,7 @@ def test_batched_fused_marlin_moe( f"topk={topk}, " f"max_tokens_per_batch={max_tokens_per_batch}" ) - torch.cuda.manual_seed(0) + set_random_seed(0) dtype = torch.bfloat16 quant_dtype = scalar_types.float4_e2m1f diff --git a/tests/kernels/moe/test_shared_fused_moe_routed_transform.py b/tests/kernels/moe/test_shared_fused_moe_routed_transform.py index 366009dce9..89e23eb0d7 100644 --- a/tests/kernels/moe/test_shared_fused_moe_routed_transform.py +++ b/tests/kernels/moe/test_shared_fused_moe_routed_transform.py @@ -15,7 +15,7 @@ from vllm.config import VllmConfig, set_current_vllm_config from vllm.forward_context import set_forward_context from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE from vllm.platforms import current_platform -from vllm.utils.torch_utils import is_torch_equal_or_newer +from vllm.utils.torch_utils import is_torch_equal_or_newer, set_random_seed class SimpleLinear(nn.Module): @@ -144,8 +144,7 @@ def test_routed_input_transform_inside_vs_outside( rocm_aiter_ops.refresh_env_variables() - torch.manual_seed(42) - torch.cuda.manual_seed(42) + set_random_seed(42) num_experts = 8 top_k = 2 diff --git a/tests/kernels/test_fused_quant_activation.py b/tests/kernels/test_fused_quant_activation.py index 2670f224d7..0696ebb8d5 100644 --- a/tests/kernels/test_fused_quant_activation.py +++ b/tests/kernels/test_fused_quant_activation.py @@ -7,6 +7,7 @@ import vllm._custom_ops as ops from tests.kernels.utils import opcheck from vllm.model_executor.layers.activation import SiluAndMul from vllm.platforms import current_platform +from vllm.utils.torch_utils import set_random_seed DTYPES = [torch.bfloat16, torch.float16] QUANT_DTYPES = [current_platform.fp8_dtype()] @@ -49,9 +50,7 @@ def test_silu_and_mul( seed: int, device: str, ) -> None: - torch.random.manual_seed(seed) - if torch.cuda.is_available(): - torch.cuda.manual_seed(seed) + set_random_seed(seed) torch.set_default_device(device) layer = SiluAndMul() diff --git a/tests/models/language/pooling/test_token_classification.py b/tests/models/language/pooling/test_token_classification.py index 42511f22f5..be71f7918e 100644 --- a/tests/models/language/pooling/test_token_classification.py +++ b/tests/models/language/pooling/test_token_classification.py @@ -1,25 +1,20 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project -import random -import numpy as np import pytest import torch from transformers import AutoModelForTokenClassification from tests.models.utils import softmax from vllm.platforms import current_platform +from vllm.utils.torch_utils import set_random_seed @pytest.fixture(autouse=True) def seed_everything(): """Seed all random number generators for reproducibility.""" seed = 0 - random.seed(seed) - np.random.seed(seed) - torch.manual_seed(seed) - if torch.cuda.is_available(): - torch.cuda.manual_seed_all(seed) + set_random_seed(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False yield diff --git a/tests/v1/e2e/spec_decode/test_lora_with_spec_decode.py b/tests/v1/e2e/spec_decode/test_lora_with_spec_decode.py index 5cbdc41232..9e000223c1 100644 --- a/tests/v1/e2e/spec_decode/test_lora_with_spec_decode.py +++ b/tests/v1/e2e/spec_decode/test_lora_with_spec_decode.py @@ -5,9 +5,6 @@ This script contains: 1. test lora with speculative decoding for batch inference """ -import random - -import numpy as np import pytest import torch @@ -15,6 +12,7 @@ from vllm import LLM, SamplingParams from vllm.distributed import cleanup_dist_env_and_memory from vllm.lora.request import LoRARequest from vllm.platforms import current_platform +from vllm.utils.torch_utils import set_random_seed LORA_TEST_PROMPT_MAP: dict[str, str] = {} @@ -63,10 +61,7 @@ def test_batch_inference_correctness( with monkeypatch.context() as m: # Disable randomness m.setenv("CUBLAS_WORKSPACE_CONFIG", ":4096:8") - torch.manual_seed(SEED) - np.random.seed(SEED) - random.seed(SEED) - torch.cuda.manual_seed_all(SEED) + set_random_seed(SEED) torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True diff --git a/tests/v1/spec_decode/test_tree_attention.py b/tests/v1/spec_decode/test_tree_attention.py index cb487acec0..1b6fa4f6f4 100644 --- a/tests/v1/spec_decode/test_tree_attention.py +++ b/tests/v1/spec_decode/test_tree_attention.py @@ -14,6 +14,7 @@ from tests.v1.attention.utils import ( ) from vllm.config import ParallelConfig, SpeculativeConfig from vllm.platforms import current_platform +from vllm.utils.torch_utils import set_random_seed from vllm.v1.attention.backend import CommonAttentionMetadata from vllm.v1.attention.backends.fa_utils import is_flash_attn_varlen_func_available from vllm.v1.attention.backends.registry import AttentionBackendEnum @@ -323,8 +324,7 @@ def forward_attention( def test_tree_attn_correctness( reference_backend: AttentionBackendEnum, ) -> None: - torch.manual_seed(42) - torch.cuda.manual_seed_all(42) + set_random_seed(42) device = "cuda" tree_attn_masks = { diff --git a/tools/pre_commit/check_torch_cuda.py b/tools/pre_commit/check_torch_cuda.py index 9bc4ab56df..08ee419515 100644 --- a/tools/pre_commit/check_torch_cuda.py +++ b/tools/pre_commit/check_torch_cuda.py @@ -9,6 +9,7 @@ import regex as re # --------------------------------------------------------------------------- # _TORCH_CUDA_PATTERNS = [ r"\btorch\.cuda\.(empty_cache|synchronize|device_count|current_device|memory_reserved|memory_allocated|max_memory_allocated|max_memory_reserved|reset_peak_memory_stats|memory_stats|set_device|device\()\b", + r"\btorch\.cuda\.(manual_seed|manual_seed_all)\b", r"\bwith\storch\.cuda\.device\b", # Calls torch.cuda.{_is_compiled/_device_count_amdsmi/_device_count_nvml} internally r"\bcuda_device_count_stateless\(\)\b", @@ -24,6 +25,14 @@ def scan_file(path: str) -> int: for match in re.finditer(pattern, content, re.MULTILINE): # Calculate line number from match position line_num = content[: match.start() + 1].count("\n") + 1 + matched_text = match.group(0) + if "manual_seed" in matched_text: + print( + f"{path}:{line_num}: " + "\033[91merror:\033[0m " + f"Found {matched_text} API call. Use set_random_seed instead." + ) + return 1 print( f"{path}:{line_num}: " "\033[91merror:\033[0m " # red color diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py index b181854722..eb4a5e1482 100644 --- a/vllm/platforms/cpu.py +++ b/vllm/platforms/cpu.py @@ -154,6 +154,10 @@ class CpuPlatform(Platform): """ torch.cpu.set_device(device) + @classmethod + def manual_seed_all(cls, seed: int) -> None: + pass + @classmethod def inference_mode(cls): return torch.no_grad() diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 66a5ee67ef..1969bf9389 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -188,6 +188,10 @@ class CudaPlatformBase(Platform): # for why and when it is needed _ = torch.zeros(1, device=device) + @classmethod + def manual_seed_all(cls, seed: int) -> None: + torch.cuda.manual_seed_all(seed) + @classmethod def get_device_capability(cls, device_id: int = 0) -> DeviceCapability | None: raise NotImplementedError diff --git a/vllm/platforms/interface.py b/vllm/platforms/interface.py index 72b714df1b..c97c9118ac 100644 --- a/vllm/platforms/interface.py +++ b/vllm/platforms/interface.py @@ -391,6 +391,11 @@ class Platform: """ raise NotImplementedError + @classmethod + def manual_seed_all(cls, seed: int) -> None: + """Set RNG seed across all devices for the current platform.""" + raise NotImplementedError + @classmethod def pre_register_and_update( cls, parser: FlexibleArgumentParser | None = None diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py index 2ba4ef3fe8..5104143780 100644 --- a/vllm/platforms/rocm.py +++ b/vllm/platforms/rocm.py @@ -605,6 +605,10 @@ class RocmPlatform(Platform): """ torch.cuda.set_device(device) + @classmethod + def manual_seed_all(cls, seed: int) -> None: + torch.cuda.manual_seed_all(seed) + @classmethod @lru_cache(maxsize=8) def get_device_capability(cls, device_id: int = 0) -> DeviceCapability | None: diff --git a/vllm/platforms/xpu.py b/vllm/platforms/xpu.py index ffc765257e..892fc0c950 100644 --- a/vllm/platforms/xpu.py +++ b/vllm/platforms/xpu.py @@ -125,6 +125,10 @@ class XPUPlatform(Platform): """ torch.xpu.set_device(device) + @classmethod + def manual_seed_all(cls, seed: int) -> None: + torch.xpu.manual_seed_all(seed) + @classmethod def get_device_capability( cls, diff --git a/vllm/utils/torch_utils.py b/vllm/utils/torch_utils.py index 94f8c096e3..a51854b696 100644 --- a/vllm/utils/torch_utils.py +++ b/vllm/utils/torch_utils.py @@ -365,8 +365,9 @@ def set_random_seed(seed: int | None) -> None: random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) - if torch.cuda.is_available(): - torch.cuda.manual_seed_all(seed) + from vllm.platforms import current_platform + + current_platform.manual_seed_all(seed) def create_kv_caches_with_random_flash(