CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-10 10:27:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : implement GEGLU_ERF and GEGLU_QUICK ops (#14445)

Browse Source
This commit is contained in:
Sigbjørn Skjæret
2025-07-03 23:07:22 +02:00
committed by GitHub
parent bee28421be
commit 28657a8229
20 changed files with 789 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
ggml/src/ggml-sycl/element_wise.cpp
View File

@@ -383,6 +383,24 @@ static void gated_op_fused_swiglu(const T * x, const T * g, T * dst, const uint6
}
}
template<typename T>
static void gated_op_fused_geglu_erf(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
const int64_t j0 = (i / n) * o0 + (i % n);
const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
dst[i] = op_gelu_erf(x[j0]) * g[j1];
}
}
template<typename T>
static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
const int64_t j0 = (i / n) * o0 + (i % n);
const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
dst[i] = op_gelu_quick(x[j0]) * g[j1];
}
}
namespace ggml_sycl_detail {
static void acc_f32_sycl(const float *x, const float *y, float *dst,
const int n_elements, const int ne10, const int ne11,
@@ -978,6 +996,28 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
});
}
static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(main_stream,
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
});
}
static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(main_stream,
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
});
}
void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
@@ -1118,3 +1158,13 @@ void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_swiglu(ctx, dst);
}
void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_geglu_erf(ctx, dst);
}
void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_geglu_quick(ctx, dst);
}

2
ggml/src/ggml-sycl/element_wise.hpp
View File

@@ -80,5 +80,7 @@ void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
#endif // GGML_SYCL_ELEMENTWISE_HPP

8
ggml/src/ggml-sycl/ggml-sycl.cpp
View File

@@ -3687,6 +3687,12 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
case GGML_GLU_OP_SWIGLU:
ggml_sycl_swiglu(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_ERF:
ggml_sycl_geglu_erf(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_QUICK:
ggml_sycl_geglu_quick(ctx, dst);
break;
default:
return false;
}
@@ -4232,6 +4238,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous_1(op->src[0]);
default:
return false;