diff options
Diffstat (limited to 'ggml-kompute.cpp')
| -rw-r--r-- | ggml-kompute.cpp | 12 |
1 files changed, 9 insertions, 3 deletions
diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp index 9a469821..3f033d58 100644 --- a/ggml-kompute.cpp +++ b/ggml-kompute.cpp @@ -1559,12 +1559,18 @@ 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; -#pragma message("TODO: add ggml_vk_soft_max() F16/F32 src1 and src2 support") + 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); - GGML_ASSERT(src2 == nullptr); + +#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; |