From 30021121d3711a6b70df31115ab8252582a053d0 Mon Sep 17 00:00:00 2001 From: Ryan Spring Date: Thu, 31 Oct 2024 22:05:04 -0700 Subject: [PATCH] guard tests --- tests/cpp/test_matmul.cpp | 6 ++++++ tests/cpp/test_multidevice_matmul.cpp | 2 +- 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/tests/cpp/test_matmul.cpp b/tests/cpp/test_matmul.cpp index c3599556e1b..c84d3a1f98e 100644 --- a/tests/cpp/test_matmul.cpp +++ b/tests/cpp/test_matmul.cpp @@ -3154,6 +3154,8 @@ TEST_P(MatmulTestWithLayout, AmpereMatmulBatchSplitKBias) { // Avoid lowering error https://github.com/NVIDIA/Fuser/issues/1808 TEST_F(MatmulTest, ReproIssue1808) { + NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(8, 0, 9, 0); + // Keep multiples of 8 to keep vectorizable. int M = 504, N = 136, K = 248; @@ -3369,6 +3371,8 @@ TEST_P(MatmulTestWithLayout, MisalignedVectorization) { // Matmul test with multiple M and N dimensions that are consecutive TEST_F(MatmulTest, MultipleConsecutiveDims) { + NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(8, 0, 9, 0); + int M1 = 126, M2 = 4, N1 = 68, N2 = 2, K = 248; Fusion fusion; @@ -3559,6 +3563,8 @@ TEST_F(MatmulTest, DISABLED_MultipleNonConsecutiveNDims) { // The batch dim is parallelized, so M1 and M2 are consecutive in shared // memory. TEST_F(MatmulTest, MultipleMDimsBatch) { + NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(7, 5, 9, 0); + int Batch = 2, M1 = 126, N = 136, M2 = 4, K = 248; Fusion fusion; diff --git a/tests/cpp/test_multidevice_matmul.cpp b/tests/cpp/test_multidevice_matmul.cpp index 3032db30b94..dde422fbc2c 100644 --- a/tests/cpp/test_multidevice_matmul.cpp +++ b/tests/cpp/test_multidevice_matmul.cpp @@ -62,7 +62,7 @@ TEST_F(DistributedMatmulTest, MulSum_LayoutTN_NoComms) { // MmaLayout::TN A(T), B(N), C(T) // A and C are sharded on dimension M // Tests local matmul with no communication - NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(8, 0, 10, 0); + NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(8, 0, 9, 0); auto fusion = std::make_unique(); FusionGuard fg(fusion.get()); auto mesh = DeviceMesh::createForNumDevices(num_devices_);