wmantly/llama.cpp-xrip
1
0 Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Aman Gupta authored and GitHub committed 2026-10-21 16:32:29 +08:00
1 parent ea9e3446b5
commit eb182d85c0
2 files changed
+8 -1
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/llama-arch.cpp
+1 -1
View File
@@ -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}},
src/llama-model-loader.cpp
+7
View File
@@ -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]);