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-cuda/ggml-cuda.cu
+15 -5
View File
@@ -1744,7 +1744,7 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
return false;
}
static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
static constexpr std::array<ggml_glu_op, 4> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI, GGML_GLU_OP_SWIGLU_CLAMP };
if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
return false;
@@ -2203,6 +2203,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_GLU_OP_GEGLU_QUICK:
ggml_cuda_op_geglu_quick(ctx, dst);
break;
case GGML_GLU_OP_SWIGLU_CLAMP:
ggml_cuda_op_swiglu_clamp(ctx, dst);
break;
default:
return false;
}
@@ -3595,6 +3598,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
fusion_data.x_scale = up_scale;
fusion_data.gate_scale = gate_scale;
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
@@ -3688,6 +3692,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
fusion_data.x_scale = up_scale;
fusion_data.gate_scale = gate_scale;
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
@@ -3744,6 +3749,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
fusion_data.x_bias = up_bias_tensor;
fusion_data.gate_bias = gate_bias_tensor;
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
fused_mul_mat_vec = true;
@@ -3757,6 +3763,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
fusion_data.x_bias = up_bias_tensor;
fusion_data.gate_bias = gate_bias_tensor;
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
fused_mul_mat_vec = true;
@@ -3781,8 +3788,9 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
ggml_cuda_mm_fusion_args_host fusion_data{};
fusion_data.gate = gate->src[0];
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.gate = gate->src[0];
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
fused_mul_mat_vec = true;
@@ -3792,8 +3800,9 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
ggml_cuda_mm_fusion_args_host fusion_data{};
fusion_data.gate = gate->src[0];
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.gate = gate->src[0];
fusion_data.glu_op = ggml_get_glu_op(glu);
fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
fused_mul_mat_vec = true;
@@ -4919,6 +4928,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
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;