ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32
| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
| qwen3moe 235B.A22B Q3_K - Medium |  96.99 GiB |   235.09 B | ROCm       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |         29.46 ± 0.00 |
:0:rocdevice.cpp            :3588: 50932421658 us:  Callback: Queue 0x7f8e6a000000 aborting with error : HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to access memory beyond the largest legal address. code: 0x29
Kernel Name: _ZL18flash_attn_ext_vecILi128ELi1EL9ggml_type1ELS0_1ELb0EEvPKcS2_S2_S2_S2_PKiPfP15HIP_vector_typeIfLj2EEffffjfiiiiiiiiiiiiiliiliiiiil
VGPU=0x94e06a0 SWq=0x7f8e6cbea000, HWq=0x7f8e6a000000, id=2
	Dispatch Header =0xb02 (type=2, barrier=1, acquire=1, release=1), setup=0
	grid=[32, 68, 64], workgroup=[32, 4, 1]
	private_seg_size=0, group_seg_size=4352
	kernel_obj=0x7f8e6a78f180, kernarg_address=0x0x7f738bd49400
	completion_signal=0x0, correlation_id=0
	rptr=1368490, wptr=1369554
 /opt/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:89: ROCm error
/usr/local/lib64/libggml-base.so.0(+0x3565) [0x7f8e79498565]
/usr/local/lib64/libggml-base.so.0(ggml_print_backtrace+0x1eb) [0x7f8e7949892b]
/usr/local/lib64/libggml-base.so.0(ggml_abort+0x11f) [0x7f8e79498aaf]
/usr/local/lib64/libggml-hip.so.0(+0x2812fb2) [0x7f8e7bd67fb2]
/usr/local/lib64/libggml-hip.so.0(+0x2818004) [0x7f8e7bd6d004]
/usr/local/lib64/libggml-base.so.0(ggml_backend_sched_synchronize+0x2e) [0x7f8e794af8ce]
/usr/local/lib64/libllama.so.0(_ZN13llama_context11synchronizeEv+0x10) [0x7f8e7c40f950]
/usr/local/bin/llama-bench() [0x408242]
/lib64/libc.so.6(+0x35b5) [0x7f8e78e2e5b5]
/lib64/libc.so.6(__libc_start_main+0x88) [0x7f8e78e2e668]
/usr/local/bin/llama-bench() [0x409255]
✖ ! [rocm-7alpha-rocwmma] Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL-00001-of-00003__hblt0__fa1 __longctx32768 failed (exit 0)