Open junliume opened 11 months ago
i have this same error
[ 18%] Building CXX object library/src/tensor_operation_instance/gpu/gemm/CMakeFiles/device_gemm_instance.dir/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp.o In file included from /tmp/ROCm-SBo/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp:9: In file included from /tmp/ROCm-SBo/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp:14: In file included from /tmp/ROCm-SBo/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp:13: In file included from /tmp/ROCm-SBo/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp:9: /tmp/ROCm-SBo/composable_kernel/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp:453:81: error: use of undeclared identifier 'f8_t' 453 | is_same<BaseType, int8_t>::value || is_same<BaseType, f8_t>::value, | ^ 1 error generated when compiling for gfx900.
cmake ... -DDTYPES="fp32;fp16;bf16" ...
resolve this by adding fp8;bf8 in my dtypes?
Yes, since we've added a few mixed-type kernels, we cannot decouple the fp16 and fp8 types.
[Reproduce]
hence
-DDTYPES="fp16;fp32;bf16"
, then[Observation]
Since fp8;bf8 was not included in the cmake command, resulting in the instances partially not protected. Hence