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

Fixed CUDA RoPE

Browse source
This commit is contained in:
JohannesGaessler 2023-06-07 11:47:52 +02:00
parent 74d4cfa343
commit 8c6bd319db
2 changed files with 8 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml-cuda.cu
View file

@ -1783,7 +1783,7 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) {
void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) { void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
if (tensor->src0 != nullptr && tensor->src0->op == GGML_OP_RESHAPE) { if (tensor->src0 != nullptr && tensor->src0->op == GGML_OP_RESHAPE) {
ggml_cuda_assign_buffers(tensor); ggml_cuda_assign_buffers(tensor->src0);
} }
const size_t size = ggml_nbytes(tensor); const size_t size = ggml_nbytes(tensor);
@ -1800,8 +1800,7 @@ void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
CUDA_CHECK(cudaSetDevice(g_main_device)); CUDA_CHECK(cudaSetDevice(g_main_device));
if (inplace && tensor->src0->backend == GGML_BACKEND_GPU) { if (inplace && tensor->src0->backend == GGML_BACKEND_GPU) {
struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src0->extra; struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src0->extra;
extra->data_device[g_main_device] = src0_extra->data_device; extra->data_device[g_main_device] = src0_extra->data_device[g_main_device];
GGML_ASSERT(false);
} else { } else {
char * data = (char *) g_scratch_buffer; char * data = (char *) g_scratch_buffer;
if (data == nullptr) { if (data == nullptr) {

8
llama.cpp
View file

@ -1366,17 +1366,21 @@ static bool llama_eval_internal(
{ {
// compute Q and K and RoPE them // compute Q and K and RoPE them
struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
// offload_func(tmpq); offload_func(tmpq);
ggml_set_name(tmpq, "tmpq"); ggml_set_name(tmpq, "tmpq");
struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
// offload_func(tmpk); offload_func(tmpk);
ggml_set_name(tmpk, "tmpk"); ggml_set_name(tmpk, "tmpk");
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0); struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
offload_func(Kcur);
Kcur->backend = GGML_BACKEND_CPU;
ggml_set_name(Kcur, "Kcur"); ggml_set_name(Kcur, "Kcur");
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0); struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
offload_func(Qcur);
Qcur->backend = GGML_BACKEND_CPU;
ggml_set_name(Qcur, "Qcur"); ggml_set_name(Qcur, "Qcur");
// store key and value to memory // store key and value to memory