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

vulkan : add dev notes

Browse source
This commit is contained in:
Georgi Gerganov 2024-05-10 15:56:25 +03:00
parent 536983b1ad
commit 397b1f8f9d
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 11 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-kompute.cpp
View file

@ -1559,12 +1559,19 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
case GGML_OP_SOFT_MAX:
{
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
float max_bias;
memcpy(&scale, (float *)dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
#pragma message("TODO: add ALiBi support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192")
GGML_ASSERT(max_bias == 0.0f);
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: