ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix bf16 additions on CUDA arch < Ampere (#1164)

Browse Source 
* Fix bf16 additions on CUDA arch < Ampere

* Prevent using NCCL if graph reduce type is bf16 and arch < AMPERE
This commit is contained in:
Kawrakow
2026-01-19 12:27:52 +02:00
committed by GitHub
parent 0c0b6e4b8b
commit 6a5c180be9
2 changed files with 25 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/ggml-cuda/binbcast.cu
View File

@@ -336,7 +336,15 @@ static __global__ void k_add_same(int64_t nelem, const data_t * x, const data_t
if (i >= nelem) {
return;
}
if constexpr (std::is_same_v<data_t, nv_bfloat16>) {
#if __CUDA_ARCH__ >= CC_AMPERE
z[i] = x[i] + y[i];
#else
z[i] = __float2bfloat16((float)x[i] + (float)y[i]);
#endif
} else {
z[i] = x[i] + y[i];
}
}
template <int block_size>
@@ -373,20 +381,6 @@ static __global__ void k_add_same_q8_0(int nelem, const block_q8_0 * x, const fl
}
}
//static __global__ void k_add_same_q8_0(const block_q8_0 * x, const block_q8_0 * y, block_q8_0 * z) {
// int ib = blockIdx.x;
// int iq = threadIdx.x;
// float s = (float)x[ib].d * x[ib].qs[iq] + (float)y[ib].d * y[ib].qs[iq];
// float as = fabsf(s);
// as = warp_reduce_max(as);
// float d = as / 127;
// float id = d > 0 ? 1/d : 0;
// z[ib].qs[iq] = roundf(s * id);
// if (threadIdx.x == 0) {
// z[ib].d = (half)d;
// }
//}
void ggml_op_add_same_type(ggml_backend_cuda_context & ctx, enum ggml_type type, size_t nelem,
const void * x, const void * y, void * z) {
constexpr int kBlockSize = 256;

16
ggml/src/ggml-cuda/reduce.cu
View File

@@ -14,7 +14,15 @@ template <typename T, int block_size>
static __global__ void k_add(int nelem, const T * __restrict__ src, T * __restrict__ dst) {
int i = blockIdx.x*block_size + threadIdx.x;
if (i >= nelem) return;
if constexpr (std::is_same_v<T, nv_bfloat16>) {
#if __CUDA_ARCH__ >= CC_AMPERE
dst[i] += src[i];
#else
dst[i] = __float2bfloat16((float)src[i] + (float)dst[i]);
#endif
} else {
dst[i] += src[i];
}
}
template <int block_size>
@@ -130,7 +138,13 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
// It does not work at all if not all GPUs participate in the reduce op, and we
// get suboptimal prompt processing performance when we have more than 2 GPUs.
// Hence, if enabled, we use NCCL only for the cases where it works and performs well.
if (info.have_nccl && dst->type != GGML_TYPE_Q8_0 && nhave == nreduce && (nhave == 2 || dst->ne[1] < 32)) {
#if __CUDA_ARCH__ >= CC_AMPERE
constexpr bool bf16_supported = true;
#else
constexpr bool bf16_supported = false;
#endif
if (info.have_nccl && dst->type != GGML_TYPE_Q8_0 && nhave == nreduce && (nhave == 2 || dst->ne[1] < 32) &&
(dst->type != GGML_TYPE_BF16 || bf16_supported)) {
GGML_ASSERT(info.have_nccl);
GGML_ASSERT(info.device_count == nreduce);
auto data_type = dst->type == GGML_TYPE_F32 ? ncclFloat : dst->type == GGML_TYPE_BF16 ? ncclBfloat16 : ncclHalf;