test

Author: ggerganov

ggml/src/ggml-cuda/convert.cu

@@ -580,7 +580,11 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
     const src_t * x = (const src_t *) vx;
 
     if constexpr (std::is_same_v<src_t, nv_bfloat16>) {
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+        y[i] = (float)x[i];
+#else
         y[i] = __bfloat162float(x[i]);
+#endif
     } else if constexpr (std::is_same_v<dst_t, nv_bfloat16> && std::is_same_v<src_t, half>) {
         y[i] = (float)x[i];
     } else {

ggml/src/ggml-cuda/vendors/hip.h

@@ -20,6 +20,7 @@
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_TF32_TENSOR_OP_MATH 0
 #define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_16BF HIPBLAS_R_16BF
 #define CUDA_R_32F  HIPBLAS_R_32F
 #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED hipDeviceAttributeVirtualMemoryManagementSupported
 #define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED hipMemAllocationGranularityRecommended

ggml/src/ggml-cuda/vendors/musa.h

@@ -15,6 +15,7 @@
 #define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
 #define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
 #define CUDA_R_16F  MUSA_R_16F
+#define CUDA_R_16BF MUSA_R_16BF
 #define CUDA_R_32F  MUSA_R_32F
 #define cublasComputeType_t cudaDataType_t
 #define cublasCreate mublasCreate