diff --git a/aten/src/ATen/cuda/detail/CUDAHooks.cpp b/aten/src/ATen/cuda/detail/CUDAHooks.cpp index 03a3a97525a43..00a9f25292a5a 100644 --- a/aten/src/ATen/cuda/detail/CUDAHooks.cpp +++ b/aten/src/ATen/cuda/detail/CUDAHooks.cpp @@ -554,8 +554,11 @@ bool CUDAHooks::isGPUArch(const std::vector& archs, DeviceIndex dev const std::vector& CUDAHooks::getHipblasltPreferredArchs() const { static const std::vector archs = { "gfx90a", "gfx942", -#if ROCM_VERSION >= 60400 - "gfx1200", "gfx1201", +#if ROCM_VERSION >= 60300 + "gfx1100", "gfx1101", "gfx1102", "gfx1200", "gfx1201", +#endif +#if ROCM_VERSION >= 60402 + "gfx1150", "gfx1151", #endif #if ROCM_VERSION >= 70000 "gfx950" @@ -570,8 +573,11 @@ const std::vector& CUDAHooks::getHipblasltSupportedArchs() const { #if ROCM_VERSION >= 60300 "gfx1100", "gfx1101", "gfx1103", "gfx1200", "gfx1201", "gfx908", #endif +#if ROCM_VERSION >= 60402 + "gfx1150", "gfx1151", +#endif #if ROCM_VERSION >= 70000 - "gfx950", "gfx1150", "gfx1151", + "gfx950", #endif #if ROCM_VERSION >= 70200 "gfx1250" diff --git a/test/test_cuda.py b/test/test_cuda.py index df9bdd5b0be11..e031d2a62a282 100644 --- a/test/test_cuda.py +++ b/test/test_cuda.py @@ -629,7 +629,12 @@ def _check_default(): gcn_arch = str( torch.cuda.get_device_properties(0).gcnArchName.split(":", 1)[0] ) - if gcn_arch in ["gfx90a", "gfx942", "gfx950", "gfx1200", "gfx1201"]: + if gcn_arch in [ + "gfx90a", "gfx942", "gfx950", + "gfx1100", "gfx1101", "gfx1102", + "gfx1150", "gfx1151", + "gfx1200", "gfx1201", + ]: self.assertTrue(default == torch._C._BlasBackend.Cublaslt) else: self.assertTrue(default == torch._C._BlasBackend.Cublas)