Recently I was using tvm for FP16 inference, using rocm5.1.0+miopen, the code is as follows, when adding mod = tvm.relay.transform.ToMixedPrecision(mixed_precision_type=‘float16’, missing_op_mode=1)(mod), and target = “rocm -libs= miopen, rocblas”, the following error is reported:
Memory access fault by GPU node-4 (Agent handle: 0x33ff6d0) on address 0x7f765cc02000. Reason: Page not present or supervisor privilege. Aborted (core dumped)。
When target = “rocm” and does not add miopen, inference can be performed normally. Can you help me analyze and solve this problem? Thank you so much.