Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/ISSUE_TEMPLATE/010-bug-compilation.yml
#	.github/ISSUE_TEMPLATE/011-bug-results.yml
#	.github/ISSUE_TEMPLATE/019-bug-misc.yml
#	.github/workflows/build.yml
#	.github/workflows/docker.yml
#	CMakeLists.txt
#	Makefile
#	Package.swift
#	examples/CMakeLists.txt
#	examples/eval-callback/CMakeLists.txt
#	examples/llama-bench/llama-bench.cpp
#	examples/server/README.md
#	examples/server/server.cpp
#	examples/simple-chat/simple-chat.cpp
#	examples/simple/simple.cpp
#	examples/speculative-simple/speculative-simple.cpp
#	examples/speculative/speculative.cpp
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-amx/CMakeLists.txt
#	ggml/src/ggml-blas/CMakeLists.txt
#	ggml/src/ggml-cann/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-cuda/CMakeLists.txt
#	ggml/src/ggml-hip/CMakeLists.txt
#	ggml/src/ggml-kompute/CMakeLists.txt
#	ggml/src/ggml-kompute/ggml-kompute.cpp
#	ggml/src/ggml-metal/CMakeLists.txt
#	ggml/src/ggml-musa/CMakeLists.txt
#	ggml/src/ggml-rpc/CMakeLists.txt
#	ggml/src/ggml-sycl/CMakeLists.txt
#	ggml/src/ggml-vulkan/CMakeLists.txt
#	pocs/CMakeLists.txt
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-quantize-fns.cpp
This commit is contained in:
Concedo
2024-11-26 17:01:20 +08:00
parent 5b35790343 0eb4e12bee
commit ec581b19d8
31 changed files with 1383 additions and 676 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama.cpp
+69 -56
View File
@@ -193,7 +193,7 @@ enum llm_arch {
LLM_ARCH_COMMAND_R,
LLM_ARCH_DBRX,
LLM_ARCH_OLMO,
LLM_ARCH_OLMO_1124,
LLM_ARCH_OLMO2,
LLM_ARCH_OLMOE,
LLM_ARCH_OPENELM,
LLM_ARCH_ARCTIC,
@@ -247,7 +247,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_COMMAND_R, "command-r" },
{ LLM_ARCH_DBRX, "dbrx" },
{ LLM_ARCH_OLMO, "olmo" },
{ LLM_ARCH_OLMO_1124, "olmo_1124" },
{ LLM_ARCH_OLMO2, "olmo2" },
{ LLM_ARCH_OLMOE, "olmoe" },
{ LLM_ARCH_OPENELM, "openelm" },
{ LLM_ARCH_ARCTIC, "arctic" },
@@ -1224,7 +1224,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
},
},
{
LLM_ARCH_OLMO_1124,
LLM_ARCH_OLMO2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
@@ -4888,7 +4888,9 @@ struct llama_model_loader {
mappings.reserve(files.size());
mmaps_used.reserve(files.size());
for (const auto & file : files) {
std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa()));
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
mmaps_used.emplace_back(mapping->size, 0);
if (mlock_mmaps) {
std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
@@ -5930,7 +5932,7 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_OLMO_1124:
case LLM_ARCH_OLMO2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -8722,7 +8724,7 @@ static bool llm_load_tensors(
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
} break;
case LLM_ARCH_OLMO_1124:
case LLM_ARCH_OLMO2:
{
model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -9322,7 +9324,7 @@ static bool llm_load_tensors(
ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
if (!dev) {
// FIXME: workaround for CPU backend buft having a NULL device
dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0);
dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
}
ggml_backend_dev_props props;
ggml_backend_dev_get_props(dev, &props);
@@ -14645,7 +14647,7 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_olmo_1124() {
struct ggml_cgraph * build_olmo2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
// mutable variable, needed during the last layer of the computation to skip unused tokens
@@ -16961,9 +16963,9 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_olmo();
} break;
case LLM_ARCH_OLMO_1124:
case LLM_ARCH_OLMO2:
{
result = llm.build_olmo_1124();
result = llm.build_olmo2();
} break;
case LLM_ARCH_OLMOE:
{
@@ -17607,8 +17609,9 @@ static enum ggml_status llama_graph_compute(
int n_threads,
ggml_threadpool * threadpool) {
if (lctx.backend_cpu != nullptr) {
ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu));
auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
set_threadpool_fn(lctx.backend_cpu, threadpool);
}
// set the number of threads for all the backends
@@ -19525,6 +19528,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
//
struct llama_model_params llama_model_default_params() {
struct llama_model_params result = {
/*.devices =*/ nullptr,
/*.n_gpu_layers =*/ 0,
/*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER,
/*.main_gpu =*/ 0,
@@ -19646,7 +19650,11 @@ void llama_backend_init(void) {
void llama_numa_init(enum ggml_numa_strategy numa) {
if (numa != GGML_NUMA_STRATEGY_DISABLED) {
ggml_numa_init(numa);
auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
GGML_ASSERT(dev && "CPU backend is not loaded");
auto * reg = ggml_backend_dev_backend_reg(dev);
auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init");
numa_init_fn(numa);
}
}
@@ -19737,19 +19745,24 @@ struct llama_model * llama_load_model_from_file(
}
// create list of devices to use with this model
// currently, we use all available devices
// TODO: rework API to give user more control over device selection
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
switch (ggml_backend_dev_type(dev)) {
case GGML_BACKEND_DEVICE_TYPE_CPU:
case GGML_BACKEND_DEVICE_TYPE_ACCEL:
// skip CPU backends since they are handled separately
break;
if (params.devices) {
for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
model->devices.push_back(*dev);
}
} else {
// use all available devices
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
switch (ggml_backend_dev_type(dev)) {
case GGML_BACKEND_DEVICE_TYPE_CPU:
case GGML_BACKEND_DEVICE_TYPE_ACCEL:
// skip CPU backends since they are handled separately
break;
case GGML_BACKEND_DEVICE_TYPE_GPU:
model->devices.push_back(dev);
break;
case GGML_BACKEND_DEVICE_TYPE_GPU:
model->devices.push_back(dev);
break;
}
}
}
@@ -19920,9 +19933,6 @@ struct llama_context * llama_new_context_with_model(
__func__, n_ctx_per_seq, hparams.n_ctx_train);
}
ctx->abort_callback = params.abort_callback;
ctx->abort_callback_data = params.abort_callback_data;
ctx->logits_all = params.logits_all;
// build worst-case graph for encoder if a model contains encoder
@@ -19971,7 +19981,7 @@ struct llama_context * llama_new_context_with_model(
}
// add CPU backend
ctx->backend_cpu = ggml_backend_cpu_init();
ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
if (ctx->backend_cpu == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
llama_free(ctx);
@@ -19991,6 +20001,8 @@ struct llama_context * llama_new_context_with_model(
}
}
llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data);
if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
llama_free(ctx);
@@ -20036,7 +20048,8 @@ struct llama_context * llama_new_context_with_model(
std::vector<ggml_backend_t> backend_ptrs;
for (auto & backend : ctx->backends) {
auto * buft = ggml_backend_get_default_buffer_type(backend.get());
if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) {
auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) {
// use the host buffer of the first device CPU for faster transfer of the intermediate state
auto * dev = model->devices[0];
auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
@@ -20064,7 +20077,8 @@ struct llama_context * llama_new_context_with_model(
// pipeline parallelism requires support for async compute and events in all devices
if (pipeline_parallel) {
for (auto & backend : ctx->backends) {
if (ggml_backend_is_cpu(backend.get())) {
auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
// ignore CPU backend
continue;
}
@@ -20238,7 +20252,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_QWEN:
case LLM_ARCH_QWEN2:
case LLM_ARCH_QWEN2MOE:
case LLM_ARCH_OLMO_1124:
case LLM_ARCH_OLMO2:
case LLM_ARCH_OLMOE:
case LLM_ARCH_PHI2:
case LLM_ARCH_PHI3:
@@ -21628,6 +21642,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) {
void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
for (auto & backend : ctx->backends) {
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
if (set_abort_callback_fn) {
set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data);
}
}
}
void llama_set_embeddings(struct llama_context * ctx, bool embeddings) {
@@ -22369,32 +22391,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int
}
const char * llama_print_system_info(void) {
ggml_cpu_init(); // some ARM features are detected at runtime
static std::string s;
s = "";
s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | ";
s += "AVX_VNNI = " + std::to_string(ggml_cpu_has_avx_vnni()) + " | ";
s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | ";
s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | ";
s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
s += "AMX_INT8 = " + std::to_string(ggml_cpu_has_amx_int8()) + " | ";
s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | ";
s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | ";
s += "SVE = " + std::to_string(ggml_cpu_has_sve()) + " | ";
s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | ";
s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | ";
s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | ";
s += "RISCV_VECT = " + std::to_string(ggml_cpu_has_riscv_v()) + " | ";
s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | ";
s += "SSSE3 = " + std::to_string(ggml_cpu_has_ssse3()) + " | ";
s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | ";
s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
s += "LLAMAFILE = " + std::to_string(ggml_cpu_has_llamafile()) + " | ";
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
auto * reg = ggml_backend_reg_get(i);
auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
if (get_features_fn) {
ggml_backend_feature * features = get_features_fn(reg);
s += ggml_backend_reg_name(reg);
s += " : ";
for (; features->name; features++) {
s += features->name;
s += " = ";
s += features->value;
s += " | ";
}
}
}
return s.c_str();
}