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

fix cuda

Browse source
This commit is contained in:
slaren 2024-03-30 00:10:02 +01:00
parent 325e5efa0d
commit 26c09adce6
2 changed files with 12 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-cuda.cu
View file

@ -401,10 +401,8 @@ GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t
GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
if (tensor->view_src != NULL && tensor->view_offs == 0) {
if (tensor->view_src != NULL) {
assert(tensor->view_src->buffer->buft == buffer->buft);
tensor->backend = tensor->view_src->backend;
tensor->extra = tensor->view_src->extra;
return;
}
@ -1965,6 +1963,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * ids = dst->src[2];
GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
@ -1973,7 +1972,6 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
const size_t nb11 = src1->nb[1];
const size_t nb1 = dst->nb[1];
const struct ggml_tensor * ids = dst->src[2];
const int32_t id = ((int32_t *) dst->op_params)[0];
const int32_t n_as = src0->ne[2];
@ -1991,7 +1989,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
char * dst_original = (char *) dst->data;
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[3] = src0->nb[2];
if (src1->ne[1] == 1) {
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {

13
llama.cpp
View file

@ -4466,17 +4466,20 @@ static bool llm_load_tensors(
// hack to merge tensors, need to clean this up
// merged tensors
ggml_type type = ml.get_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type, n_embd, n_ff, hparams.n_expert);
layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type, n_ff, n_embd, hparams.n_expert);
layer.ffn_up_exps = ggml_new_tensor_3d(ctx_split, type, n_embd, n_ff, hparams.n_expert);
ggml_type type_gate = ml.get_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
ggml_type type_down = ml.get_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).c_str())->type;
ggml_type type_up = ml.get_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, 0).c_str())->type;
layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd, n_ff, hparams.n_expert);
layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down, n_ff, n_embd, hparams.n_expert);
layer.ffn_up_exps = ggml_new_tensor_3d(ctx_split, type_up, n_embd, n_ff, hparams.n_expert);
// MoE branch
for (uint32_t x = 0; x < hparams.n_expert; ++x) {
// individual tensors as views
ggml_tensor * ffn_gate_exp = ggml_view_2d(ctx_split, layer.ffn_gate_exps, n_embd, n_ff, layer.ffn_gate_exps->nb[1], layer.ffn_gate_exps->nb[2]*x);
ggml_tensor * ffn_down_exp = ggml_view_2d(ctx_split, layer.ffn_down_exps, n_ff, n_embd, layer.ffn_down_exps->nb[1], layer.ffn_down_exps->nb[2]*x);
ggml_tensor * ffn_up_exp = ggml_view_2d(ctx_split, layer.ffn_up_exps, n_embd, n_ff, layer.ffn_up_exps->nb[1], layer.ffn_up_exps->nb[2]*x);
ggml_tensor * ffn_up_exp = ggml_view_2d(ctx_split, layer.ffn_up_exps, n_embd, n_ff, layer.ffn_up_exps->nb[1], layer.ffn_up_exps->nb[2]*x);
ggml_set_name(ffn_gate_exp, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x).c_str());
ggml_set_name(ffn_down_exp, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x).c_str());