happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

CUDA: static assert to prevent misuse of memcpy_1 (#17198)

This commit is contained in:
Johannes Gäßler 2025-11-12 23:13:55 +01:00 committed by GitHub
parent 92bb442ad9
commit 5d6838b74f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cuda/common.cuh
View File

@ -586,6 +586,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
// If dst and src point at different address spaces then they are guaranteed to not be aliased. // If dst and src point at different address spaces then they are guaranteed to not be aliased.
template <int nbytes, int alignment = 0> template <int nbytes, int alignment = 0>
static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) { static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
static_assert(
nbytes <= ggml_cuda_get_max_cpy_bytes() || alignment == 0,
"You are misusing the alignment parameter for ggml_cuda_memcpy_1. "
"The intent is for the parameter is only as a workaround if either one of the pointers is not properly aligned. "
"If you use it to do more bytes per copy than ggml_cuda_max_cpy_bytes() the reads and writes may not be coalesced. "
"Call ggml_cuda_memcpy_1 in a loop instead.");
if constexpr (alignment != 0) { if constexpr (alignment != 0) {
static_assert(nbytes % alignment == 0, "bad alignment"); static_assert(nbytes % alignment == 0, "bad alignment");
} }