vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : add TODO's for F16/F32 mask/pos support in other backends

Browse source
This commit is contained in:
Georgi Gerganov 2024-04-23 10:01:49 +03:00
parent c129369702
commit 3864eea4cb
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
3 changed files with 17 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml-kompute.cpp
View file

@ -1427,6 +1427,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
for (int i = node_start; i < node_end; ++i) {
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
struct ggml_tensor * dst = gf->nodes[i];
GGML_ASSERT(dst->data != nullptr);
@ -1559,6 +1560,12 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
{
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
#pragma message("TODO: add ggml_vk_soft_max() F16/F32 src1 and src2 support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
GGML_ASSERT(src2 == nullptr);
ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
} break;
case GGML_OP_DIAG_MASK_INF: