File size: 2,091 Bytes
25b4ce2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 |
#ifndef _compat_gemm_cuh
#define _compat_gemm_cuh
#if defined(USE_ROCM)
// For some reason this include is not present anywhere in exllama_v2 codebase, but it is required
// for symbols as hipblasHalf.
#include <hipblas/hipblas.h>
__host__ __forceinline__ hipblasStatus_t __compat_hipblasHgemm(hipblasHandle_t handle,
hipblasOperation_t transA,
hipblasOperation_t transB,
int m,
int n,
int k,
const half* alpha,
const half* AP,
int lda,
const half* BP,
int ldb,
const half* beta,
half* CP,
int ldc) {
return hipblasHgemm(handle, transA, transB, m, n, k,
reinterpret_cast<const hipblasHalf *>(alpha),
reinterpret_cast<const hipblasHalf *>(AP), lda,
reinterpret_cast<const hipblasHalf *>(BP), ldb,
reinterpret_cast<const hipblasHalf *>(beta),
reinterpret_cast<hipblasHalf *>(CP), ldc);
}
#define hipblasHgemm __compat_hipblasHgemm
// Previous version of PyTorch were converting to rocBLAS instead of hipBLAS.
#define rocblas_operation_none HIPBLAS_OP_N
#define rocblas_hgemm __compat_hipblasHgemm
#endif
#endif
|