ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: FA support for Deepseek (Ampere or newer) (#13306)

Browse source 
* CUDA: FA support for Deepseek (Ampere or newer)

* do loop unrolling via C++ template
This commit is contained in:
Johannes Gäßler 2025-05-09 13:34:58 +02:00 • committed by GitHub
parent 27ebfcacba
commit 0cf6725e9f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
33 changed files with 852 additions and 547 deletions
Download patch file Download diff file Expand all files Collapse all files

19
ggml/src/ggml-cuda/common.cuh
View file

@ -296,6 +296,25 @@ static __device__ void no_device_code(
#define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
#endif // __CUDA_ARCH__
// The compiler is always able to unroll loops if they contain continue expressions.
// In such cases loop unrolling can still be achieved via recursion:
template <int n>
struct ggml_cuda_unroll {
template <typename Func, typename... Args>
__device__ void operator()(const Func & f, Args... args) const {
f(n - 1, args...);
ggml_cuda_unroll<n - 1>{}(f, args...);
}
};
template <>
struct ggml_cuda_unroll<1> {
template <typename Func, typename... Args>
__device__ void operator()(const Func & f, Args... args) const {
f(0, args...);
}
};
template<int width = WARP_SIZE>
static __device__ __forceinline__ int warp_reduce_sum(int x) {
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE