Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml: add SWIGLU_CLAMP (#27930)

Browse Source 
* ggml: add SWIGLU_CLAMP

* add vulkan shader
This commit is contained in:
Aman Gupta
2026-08-30 23:00:02 +08:00
committed by GitHub
parent 2578138397
commit 0190529ec4
40 changed files with 768 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/element_wise.cpp
+101
View File
@@ -1132,6 +1132,102 @@ void ggml_sycl_op_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
swiglu_oai_sycl(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), alpha, limit, stream);
}
template <typename T>
static void swiglu_clamp_kernel(const T * gate,
const T * up,
T * dst,
const int64_t k,
const int64_t n,
const int64_t o0,
const int64_t o1,
float limit,
sycl::nd_item<3> item_ct1) {
const int64_t i = int64_t(item_ct1.get_local_range(2)) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
if (i >= k) {
return;
}
const int64_t j0 = (i / n) * o0 + (i % n);
const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
const float gate_value = sycl::fmin((float) gate[j0], limit);
const float up_value = sycl::fmax(sycl::fmin((float) up[j1], limit), -limit);
dst[i] = (T) (gate_value / (1.0f + sycl::native::exp(-gate_value)) * up_value);
}
template <typename T>
static void swiglu_clamp_sycl(const T * gate,
const T * up,
T * dst,
const int64_t k,
const int64_t n,
const int64_t o0,
const int64_t o1,
float limit,
dpct::queue_ptr stream) {
const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
[=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
swiglu_clamp_kernel(gate, up, dst, k, n, o0, o1, limit, item_ct1);
});
}
static void ggml_sycl_op_swiglu_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
void * src0_d = src0->data;
void * src1_d = src1 ? src1->data : src0->data;
const int64_t src0_o = src0->nb[1];
const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
void * dst_d = dst->data;
const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
dpct::queue_ptr stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
GGML_ASSERT(src1->ne[0] == nc);
GGML_ASSERT(src0->type == src1->type);
}
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float limit = ggml_get_op_params_f32(dst, 3);
if (src0->type == GGML_TYPE_F16) {
sycl::half * src0_p = (sycl::half *) src0_d;
sycl::half * src1_p = (sycl::half *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
swiglu_clamp_sycl(src0_p, src1_p, (sycl::half *) dst_d, ggml_nelements(dst), nc, src0_o / sizeof(sycl::half),
src1_o / sizeof(sycl::half), limit, stream);
} else {
float * src0_p = (float *) src0_d;
float * src1_p = (float *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
swiglu_clamp_sycl(src0_p, src1_p, (float *) dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float),
src1_o / sizeof(float), limit, stream);
}
}
static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_gelu_erf(x);
@@ -1295,6 +1391,11 @@ void ggml_sycl_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_swiglu_oai(ctx, dst);
}
void ggml_sycl_swiglu_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_swiglu_clamp(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);
ggml/src/ggml-sycl/element_wise.hpp
+1
View File
@@ -77,6 +77,7 @@ void ggml_sycl_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_swiglu_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
void ggml_sycl_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
ggml/src/ggml-sycl/ggml-sycl.cpp
+4
View File
@@ -5373,6 +5373,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
case GGML_GLU_OP_SWIGLU_OAI:
ggml_sycl_swiglu_oai(ctx, dst);
break;
case GGML_GLU_OP_SWIGLU_CLAMP:
ggml_sycl_swiglu_clamp(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_ERF:
ggml_sycl_geglu_erf(ctx, dst);
break;
@@ -6133,6 +6136,7 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
case GGML_GLU_OP_SWIGLU_CLAMP:
return ggml_is_contiguous_1(op->src[0]);
default:
return false;