From 3864eea4cbf4d224d8ce798e7c537838048f540a Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 23 Apr 2024 10:01:49 +0300 Subject: [PATCH] ggml : add TODO's for F16/F32 mask/pos support in other backends --- ggml-kompute.cpp | 7 +++++++ ggml-sycl.cpp | 6 +++++- ggml-vulkan.cpp | 5 +++++ 3 files changed, 17 insertions(+), 1 deletion(-) diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp index 407062e6f..9a469821d 100644 --- a/ggml-kompute.cpp +++ b/ggml-kompute.cpp @@ -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: diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index a9b310243..f8ed55eb8 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -14738,7 +14738,12 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0, GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + const ggml_tensor * src2 = dst->src[2]; + +#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 and src2 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional const int64_t ne00 = src0->ne[0]; const int64_t nrows_x = ggml_nrows(src0); @@ -14754,7 +14759,6 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0, float * src2_dd = nullptr; sycl_pool_alloc src2_f; - ggml_tensor * src2 = dst->src[2]; const bool use_src2 = src2 != nullptr; if (use_src2) { diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp index 1736ab736..f712cdd5a 100644 --- a/ggml-vulkan.cpp +++ b/ggml-vulkan.cpp @@ -3178,6 +3178,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_SOFT_MAX: +#pragma message("TODO: add ggml_vk_soft_max() F16 src1 and src2 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); + if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && (src2 == nullptr || src2->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_soft_max_f32; }