Problem Description

Calling hipblasLtMatmul on a gfx1200 GPU (AMD Radeon RX 9060 XT) causes a hard segfault (Exception Code: 0xC0000005) inside torch_hip.dll at at::cuda::_int_mm. The library does not return HIPBLAS_STATUS_NOT_SUPPORTED or any other graceful error - it crashes the process entirely, bypassing PyTorch's cublas fallback path at HIPBlas.cpp:602.

Operating System

Windows 11

CPU

Intel Core i5

GPU

AMD Radeon RX 9060 XT

ROCm Version

7.12.0a20260304

ROCm Component

hipblaslt

PyTorch Version

2.12.0a0+rocm7.12.0a20260304

Steps to Reproduce

import torch

print(torch.cuda.get_device_name(0))  # AMD Radeon RX 9060 XT
print(torch.__version__) 

a = torch.randint(-128, 127, (1024, 1024), device="cuda", dtype=torch.int8)
b = torch.randint(-128, 127, (1024, 1024), device="cuda", dtype=torch.int8)

torch._int_mm(a, b)  # segfaults with 0xC0000005

Additional Information

Crash output:

Exception Code: 0xC0000005
 #1 at::cuda::_int_mm() (torch_hip.dll+0x2646a6d)
 #3 torch::autograd::autogradNotImplementedFallback (torch_cpu.dll+0x4d42d55)
 #5 at::_ops::_int_mm::call() (torch_cpu.dll+0x125c9f6)

Expected Behavior

hipblasLt should return HIPBLAS_STATUS_NOT_SUPPORTED for unsupported architectures, allowing the caller to fall back gracefully to cublas.