huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 10:38:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

This is duplicated code on Cuda 12.2.

Browse Source 
Without it we can compile for 52 (but I get Operation Not supported
when actually trying to use those kernels).
This commit is contained in:
Nicolas Patry
2023-08-10 09:20:18 +02:00
parent de7c31bfe9
commit 66d1c093e0
1 changed files with 0 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
candle-kernels/src/compatibility.cuh
View File

@ -6,24 +6,6 @@
// FIXME: the minimum compute capabilities are just guesses since the table is not specific enough
// #if __CUDA_ARCH__ < 600
// __device__ __forceinline__ __half __hmax(__half a, __half b) {
// return __float2half(fmaxf(__half2float(a), __half2float(b)));
// }
// __device__ __forceinline__ __half __hmin(__half a, __half b) {
// return __float2half(fminf(__half2float(a), __half2float(b)));
// }
// #endif
#if __CUDA_ARCH__ < 800
__device__ __forceinline__ __half __hmax_nan(__half a, __half b) {
// return __hisnan(a) ? a : (__hisnan(b) ? b : __hmax(a, b));
}
__device__ __forceinline__ __half __hmin_nan(__half a, __half b) {
// return __hisnan(a) ? a : (__hisnan(b) ? b : __hmin(a, b));
}
#endif
#if __CUDA_ARCH__ < 600
// Copied from https://docs.nvidia.com/cuda/cuda-c-programming-guide/#atomic-functions
__device__ double atomicAdd(double* address, double val) {