From 3b829cf176aa4ef357b7edd61eb9352419a8adbd Mon Sep 17 00:00:00 2001 From: stefankoncarevic Date: Thu, 20 Aug 2026 20:40:33 +0200 Subject: [PATCH] [CI/Build][ROCm] Keep the CUDA-only kernel tests out of the ROCm run (#53113) Signed-off-by: Stefan Koncarevic --- tests/kernels/test_fp32_router_gemm.py | 7 +++++-- tests/kernels/test_kimi_k3_gemm_rs.py | 6 +++++- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/tests/kernels/test_fp32_router_gemm.py b/tests/kernels/test_fp32_router_gemm.py index 571dfa077d..6384cae7cd 100644 --- a/tests/kernels/test_fp32_router_gemm.py +++ b/tests/kernels/test_fp32_router_gemm.py @@ -16,6 +16,7 @@ import pytest import torch from vllm._custom_ops import fp32_router_gemm +from vllm.platforms import current_platform # (hidden_size, num_experts) SHAPES = [(3072, 256), (6144, 128), (6144, 256)] @@ -26,8 +27,10 @@ ATOL_BF16 = 2e-2 # bf16 activation has lower precision def _requires_sm90(): - if not torch.cuda.is_available(): - pytest.skip("CUDA not available") + # ROCm reports a CUDA-like device capability (gfx950 -> (9, 5)), which would + # pass the SM90 check below for a kernel that is only built for CUDA. + if not current_platform.is_cuda(): + pytest.skip("fp32_router_gemm is built for CUDA only") major, minor = torch.cuda.get_device_capability() if major * 10 + minor < 90: pytest.skip(f"fp32_router_gemm requires SM90+, got SM{major}{minor}") diff --git a/tests/kernels/test_kimi_k3_gemm_rs.py b/tests/kernels/test_kimi_k3_gemm_rs.py index 5ca54d2ca4..d4d6347875 100644 --- a/tests/kernels/test_kimi_k3_gemm_rs.py +++ b/tests/kernels/test_kimi_k3_gemm_rs.py @@ -12,7 +12,6 @@ from vllm.distributed.parallel_state import ( init_distributed_environment, initialize_model_parallel, ) -from vllm.models.kimi_k3.nvidia.ops.cute_dsl.gemm_rs import GemmRS from vllm.platforms import current_platform from vllm.utils.network_utils import get_open_port from vllm.utils.system_utils import update_environment_variables @@ -70,6 +69,11 @@ def _assert_valid_rows_close( def _worker(local_rank: int, world_size: int, master_port: int) -> None: + # This module pulls in cute_dsl, which is unavailable off CUDA, so importing + # it at module level would fail collection. Import it where it is used, as + # `kda.py` does. + from vllm.models.kimi_k3.nvidia.ops.cute_dsl.gemm_rs import GemmRS + device = torch.device("cuda", local_rank) torch.accelerator.set_device_index(device) update_environment_variables(