vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	CMakeLists.txt
#	Makefile
#	build.zig
#	scripts/sync-ggml.sh
This commit is contained in:
Concedo 2023-10-10 00:09:46 +08:00
commit f288c6b5e3
18 changed files with 1335 additions and 363 deletions
Download patch file Download diff file Expand all files Collapse all files

48
llama.cpp
View file

@ -1330,7 +1330,11 @@ static bool llama_kv_cache_init(
cache.cells.clear();
cache.cells.resize(n_ctx);
// TODO: this should be:
// cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
// change it and test that it works
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
memset(cache.buf.data, 0, cache.buf.size);
struct ggml_init_params params;
params.mem_size = cache.buf.size;
@ -1736,7 +1740,7 @@ struct llama_model_loader {
}
}
struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend backend) {
struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend_type backend) {
if (backend != GGML_BACKEND_CPU) {
ggml_set_no_alloc(ctx, true);
}
@ -1754,7 +1758,7 @@ struct llama_model_loader {
return tensor;
}
struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, ggml_backend backend) {
struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, ggml_backend_type backend) {
struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str());
if (cur == NULL) {
@ -2305,8 +2309,8 @@ static void llm_load_tensors(
// output
{
ggml_backend backend_norm;
ggml_backend backend_output;
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@ -2341,8 +2345,8 @@ static void llm_load_tensors(
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
auto & layer = model.layers[i];
@ -2371,8 +2375,8 @@ static void llm_load_tensors(
{
model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
{
ggml_backend backend_norm;
ggml_backend backend_output;
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@ -2407,8 +2411,8 @@ static void llm_load_tensors(
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
auto & layer = model.layers[i];
@ -2441,8 +2445,8 @@ static void llm_load_tensors(
// output
{
ggml_backend backend_norm;
ggml_backend backend_output;
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@ -2479,8 +2483,8 @@ static void llm_load_tensors(
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
auto & layer = model.layers[i];
@ -2518,8 +2522,8 @@ static void llm_load_tensors(
// output
{
ggml_backend backend_norm;
ggml_backend backend_output;
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@ -2556,8 +2560,8 @@ static void llm_load_tensors(
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
auto & layer = model.layers[i];
@ -2595,8 +2599,8 @@ static void llm_load_tensors(
model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
{
ggml_backend backend_norm;
ggml_backend backend_output;
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
@ -2630,8 +2634,8 @@ static void llm_load_tensors(
const int i_gpu_start = n_layer - n_gpu_layers;
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT;
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT;
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend);