Skip to content

LLM inference in C/C++

C++ ggml

llama : place d2t by its set_rows use so it can stay on the GPU (#30912)

Aman Guptaeb182d85c0e3
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 1b954f176c..4534be22f3 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -1110,7 +1110,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_MASKED_EMBD_ORDERING, {LLM_TENSOR_LAYER_INPUT, GGML_OP_NONE}},
// eagle3
{LLM_TENSOR_FC, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
- {LLM_TENSOR_D2T, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
+ {LLM_TENSOR_D2T, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_SET_ROWS}},
// dspark
{LLM_TENSOR_DSPARK_MARKOV_W1, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_DSPARK_MARKOV_W2, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index bc93c6f9e6..9dc366fda5 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -951,6 +951,13 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 512);
op_tensor = ggml_get_rows(ctx, w, b);
} break;
+ case GGML_OP_SET_ROWS:
+ {
+ // w holds the row indices (e.g. d2t)
+ ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, w->ne[0]);
+ ggml_tensor * b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, w->ne[0]);
+ op_tensor = ggml_set_rows(ctx, a, b, w);
+ } break;
case GGML_OP_MUL_MAT:
{
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], 512, w->ne[2], w->ne[3]);