nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Faster Qwen3-Next PP on CUDA - optimize concat (#1275)

This commit is contained in:
Kawrakow 2026-02-16 11:46:39 +01:00 committed by GitHub
parent 97e7c091cd
commit 35c99f9f41
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 8 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cuda/concat.cu
View File

@ -266,6 +266,14 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst) && dim == 2 && dst->ne[3] > 1 && src1->ne[2] == 1) {
float * dst_d = (float *)dst->data;
float * src0_d = (float *)src0->data;
float * src1_d = (float *)src1->data;
concat_f32_cuda(src0_d, src1_d, dst_d, src0->ne[0]*src0->ne[1]*src0->ne[2], src0->ne[3], 1, dst->ne[0]*dst->ne[1]*dst->ne[2], dst->ne[3], 1, 0, stream);
return;
}
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
//if (dst->ne[1] >= 65536 || dst->ne[2] >= 65536) {
// fprintf(stderr, "%s: ne1 = %ld, ne2 = %ld exceed max. blocks when computing %s\n", __func__, dst->ne[1], dst->ne[2], dst->name);