linked command failed with exit code 1

[mariano54]

I'm trying to compile this for an AMD 6900XT. On an AMD CPU, arch linux. Pytorch works properly with ROCM.

Here are some of the link issues I'm getting:

/usr/bin/ld: CMakeFiles/ctranslate2.dir/src/ops/quantize_gpu.cu.o: in function `__float22bfloat162_rn(HIP_vector_type<float, 2u>)':
quantize_gpu.cu:(.text+0xa0): multiple definition of `__float22bfloat162_rn(HIP_vector_type<float, 2u>)'; CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:primitives.cu:(.text+0xa0): first defined here
/usr/bin/ld: CMakeFiles/ctranslate2.dir/src/ops/quantize_gpu.cu.o: in function `__high2float(__hip_bfloat162)':
quantize_gpu.cu:(.text+0x120): multiple definition of `__high2float(__hip_bfloat162)'; CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:primitives.cu:(.text+0x120): first defined here
/usr/bin/ld: CMakeFiles/ctranslate2.dir/src/ops/quantize_gpu.cu.o: in function `__low2float(__hip_bfloat162)':
quantize_gpu.cu:(.text+0x130): multiple definition of `__low2float(__hip_bfloat162)'; CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:primitives.cu:(.text+0x130): first defined here
clang++: error: linker command failed with exit code 1 (use -v to see invocation)

Would appreciate any help you can give here.

[mariano54]

I had to perform this patch to get it to compile:

https://github.com/vllm-project/vllm/pull/2790/files

[creeloper27]

I had to perform this patch to get it to compile:

https://github.com/vllm-project/vllm/pull/2790/files

Update: I uninstalled everything AMD, HIP and ROCm related and reinstalled everything using amdgpu-install (sudo amdgpu-install --usercases=dkms,graphics,hip,hiplibsdk,rocm,rocmdev) and a missing package I needed called miopen-hip-dev (sudo apt install miopen-hip-dev on Ubuntu 24.04), now the project succesfully compiles without any error! I didn't even need to edit any file, the stock 6.2 ROCm installation worked perfectly fine.

I am using ROCm 6.2 an I am having toruble getting this fork to work.

@mariano54 Are you able to explain what you did exactly by any chance?

I am also getting similar multiple definitions errors

[...]

[ 90%] Linking CXX shared library libctranslate2.so
ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x40)

ld.lld: error: too many errors emitted, stopping now (use --error-limit=0 to see all errors)
clang++: error: linker command failed with exit code 1 (use -v to see invocation)
gmake[2]: *** [CMakeFiles/ctranslate2.dir/build.make:2050: libctranslate2.so.4.1.0] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:192: CMakeFiles/ctranslate2.dir/all] Error 2
gmake: *** [Makefile:136: all] Error 2

I wanted to try and apply the patch you suggested to /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h but the definitions are slightly different on the file for me so I tried to adapt them.

/opt/rocm/include/hip/amd_detail/amd_hip_bf16.h (original file)

 102
 103 #include "amd_hip_vector_types.h"  // float2 etc
 104 #include "device_library_decls.h"  // ocml conversion functions
 105 #include "math_fwd.h"              // ocml device functions
 106
 107 #define __BF16_DEVICE__ __device__
 108 #if defined(__HIPCC_RTC__)
 109 #define __BF16_HOST_DEVICE__ __BF16_DEVICE__
 110 #else
 111 #include <algorithm>
 112 #include <climits>
 113 #include <cmath>
 114 #define __BF16_HOST_DEVICE__ __host__ __BF16_DEVICE__
 115 #endif
 116 #define __BF16_DEVICE_STATIC__ __BF16_DEVICE__ static inline
 117 #define __BF16_HOST_DEVICE_STATIC__ __BF16_HOST_DEVICE__ static inline
 118

/opt/rocm/include/hip/amd_detail/amd_hip_bf16.h (tweaked file)

 102
 103 #include "amd_hip_vector_types.h"  // float2 etc
 104 #include "device_library_decls.h"  // ocml conversion functions
 105 #include "math_fwd.h"              // ocml device functions
 106
 107 #define __BF16_DEVICE__ __device__
 108 #if defined(__HIPCC_RTC__)
 109 #define __BF16_HOST_DEVICE__ __BF16_DEVICE__ static
 110 #else
 111 #include <algorithm>
 112 #include <climits>
 113 #include <cmath>
 114 #define __BF16_HOST_DEVICE__ __host__ __BF16_DEVICE__ static inline
 115 #endif
 116 #define __BF16_DEVICE_STATIC__ __BF16_DEVICE__ static inline
 117 #define __BF16_HOST_DEVICE_STATIC__ __BF16_HOST_DEVICE__ static inline
 118

And got (same output as running with original files):

[...]

[ 90%] Linking CXX shared library libctranslate2.so
ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at concat_split_slide_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/concat_split_slide_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at alibi_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/alibi_add_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __bfloat1622float2(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__bfloat1622float2(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x40)

ld.lld: error: duplicate symbol: __double2bfloat16(double)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__double2bfloat16(double))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x60)

ld.lld: error: duplicate symbol: __float22bfloat162_rn(HIP_vector_type<float, 2u>)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float22bfloat162_rn(HIP_vector_type<float, 2u>))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0xA0)

ld.lld: error: duplicate symbol: __high2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__high2float(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x120)

ld.lld: error: duplicate symbol: __low2float(__hip_bfloat162)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__low2float(__hip_bfloat162))
>>> defined at dequantize_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/dequantize_gpu.cu.o:(.text+0x130)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at bias_add_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/bias_add_gpu.cu.o:(.text+0x0)

ld.lld: error: duplicate symbol: __float2bfloat16(float)
>>> defined at primitives.cu
>>>            CMakeFiles/ctranslate2.dir/src/cuda/primitives.cu.o:(__float2bfloat16(float))
>>> defined at gather_gpu.cu
>>>            CMakeFiles/ctranslate2.dir/src/ops/gather_gpu.cu.o:(.text+0x0)

ld.lld: error: too many errors emitted, stopping now (use --error-limit=0 to see all errors)
clang++: error: linker command failed with exit code 1 (use -v to see invocation)
gmake[2]: *** [CMakeFiles/ctranslate2.dir/build.make:2050: libctranslate2.so.4.1.0] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:192: CMakeFiles/ctranslate2.dir/all] Error 2
gmake: *** [Makefile:136: all] Error 2

As this did not resolve the error, I restored the original amd_hip_bf16.h file and instead tried tweaking amd_hip_bfloat16.h.

/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h (original file)

 31
 32 #include "host_defines.h"
 33 #if defined(__HIPCC_RTC__)
 34     #define __HOST_DEVICE__ __device__
 35 #else
 36     #define __HOST_DEVICE__ __host__ __device__
 37 #endif
 38

/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h (tweaked file)

 31
 32 #include "host_defines.h"
 33 #if defined(__HIPCC_RTC__)
 34     #define __HOST_DEVICE__ __device__ static
 35 #else
 36     #define __HOST_DEVICE__ __host__ __device__ static inline
 37 #endif
 38

And got:

[...]

In file included from /opt/rocm/include/hip/hip_bfloat16.h:37:
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:66:21: error: constructor cannot be declared 'static'
   66 |     __HOST_DEVICE__ hip_bfloat16() = default;
      |     ~~~~~~~~~~~~~~~ ^~~~~~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:69:30: error: constructor cannot be declared 'static'
   69 |     explicit __HOST_DEVICE__ hip_bfloat16(float f)
      |              ~~~~~~~~~~~~~~~ ^~~~~~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:74:30: error: constructor cannot be declared 'static'
   74 |     explicit __HOST_DEVICE__ hip_bfloat16(float f, truncate_t)
      |              ~~~~~~~~~~~~~~~ ^~~~~~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:80:38: error: static member function cannot have 'const' qualifier
   80 |     __HOST_DEVICE__ operator float() const
      |                                      ^
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:80:21: error: conversion function must be a non-static member function
   80 |     __HOST_DEVICE__ operator float() const
      |     ~~~~~~~~~~~~~~~ ^~~~~~~~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:90:35: error: overloaded 'operator=' cannot be a static member function
   90 |     __HOST_DEVICE__ hip_bfloat16 &operator=(const float& f)
      |                                   ^
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:92:8: error: invalid use of member 'data' in static member function
   92 |        data = float_to_bfloat16(f);
      |        ^~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:93:16: error: invalid use of 'this' outside of a non-static member function
   93 |        return *this;
      |                ^
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:185:16: error: cannot convert 'const hip_bfloat16' to 'float' without a conversion operator
  185 |   return os << float(bf16);
      |                ^~~~~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:200:25: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  200 |     return hip_bfloat16(float(a) + float(b));
      |                         ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:200:36: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  200 |     return hip_bfloat16(float(a) + float(b));
      |                                    ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:204:25: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  204 |     return hip_bfloat16(float(a) - float(b));
      |                         ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:204:36: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  204 |     return hip_bfloat16(float(a) - float(b));
      |                                    ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:208:25: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  208 |     return hip_bfloat16(float(a) * float(b));
      |                         ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:208:36: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  208 |     return hip_bfloat16(float(a) * float(b));
      |                                    ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:212:25: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  212 |     return hip_bfloat16(float(a) / float(b));
      |                         ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:212:36: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  212 |     return hip_bfloat16(float(a) / float(b));
      |                                    ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:216:12: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  216 |     return float(a) < float(b);
      |            ^~~~~~~~
/opt/rocm/include/hip/amd_detail/amd_hip_bfloat16.h:216:23: error: cannot convert 'hip_bfloat16' to 'float' without a conversion operator
  216 |     return float(a) < float(b);
      |                       ^~~~~~~~
fatal error: too many errors emitted, stopping now [-ferror-limit=]

[...]

gmake[1]: *** [CMakeFiles/Makefile2:192: CMakeFiles/ctranslate2.dir/all] Error 2
gmake: *** [Makefile:136: all] Error 2