vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-17 04:19:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/docker.yml
#	CMakeLists.txt
#	CONTRIBUTING.md
#	docs/android.md
#	docs/docker.md
#	examples/embedding/embedding.cpp
#	examples/imatrix/imatrix.cpp
#	examples/infill/infill.cpp
#	examples/llama-bench/llama-bench.cpp
#	examples/main/README.md
#	examples/parallel/parallel.cpp
#	examples/perplexity/perplexity.cpp
#	examples/quantize-stats/quantize-stats.cpp
#	examples/save-load-state/save-load-state.cpp
#	examples/server/README.md
#	examples/simple/CMakeLists.txt
#	examples/speculative/speculative.cpp
#	flake.lock
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-blas.cpp
#	pocs/vdot/q8dot.cpp
#	pocs/vdot/vdot.cpp
#	scripts/debug-test.sh
#	scripts/sync-ggml.last
#	src/llama.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tests/test-quantize-fns.cpp
#	tests/test-quantize-perf.cpp
#	tests/test-tokenizer-0.cpp
#	tests/test-tokenizer-1-bpe.cpp
#	tests/test-tokenizer-1-spm.cpp
This commit is contained in:
Concedo 2024-10-11 11:59:59 +08:00
commit e692a79aab
61 changed files with 2579 additions and 1949 deletions
Download patch file Download diff file Expand all files Collapse all files

238
src/llama.cpp
View file

@ -11,10 +11,6 @@
#include "ggml-alloc.h"
#include "ggml-backend.h"
#ifdef GGML_USE_RPC
# include "ggml-rpc.h"
#endif
#ifdef GGML_USE_CUDA
# include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST)
@ -29,14 +25,6 @@
# include "ggml-cann.h"
#endif
#ifdef GGML_USE_BLAS
# include "ggml-blas.h"
#endif
#ifdef GGML_USE_METAL
# include "ggml-metal.h"
#endif
// TODO: replace with ggml API call
#define QK_K 256
@ -3307,12 +3295,8 @@ struct llama_context {
std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
std::vector<ggml_backend_t> backends;
#ifdef GGML_USE_METAL
ggml_backend_t backend_metal = nullptr;
#endif
#ifdef GGML_USE_BLAS
ggml_backend_t backend_blas = nullptr;
#endif
std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
ggml_backend_t backend_cpu = nullptr;
ggml_threadpool_t threadpool = nullptr;
@ -3431,13 +3415,7 @@ struct llama_lora_adapter {
static int llama_get_device_count(const llama_model & model) {
int count = (int) model.devices.size();
#if defined(GGML_USE_RPC)
count += (int) model.rpc_servers.size();
#endif
#if defined(GGML_USE_METAL)
count += 1;
#elif defined(GGML_USE_SYCL)
#if defined(GGML_USE_SYCL)
count += ggml_backend_sycl_get_device_count();
#elif defined(GGML_USE_VULKAN)
count += ggml_backend_vk_get_device_count();
@ -3489,23 +3467,12 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int device) {
ggml_backend_buffer_type_t buft = nullptr;
#if defined(GGML_USE_RPC)
int rpc_count = (int)model.rpc_servers.size();
if (device < rpc_count) {
const char * endpoint = model.rpc_servers[device].c_str();
return ggml_backend_rpc_buffer_type(endpoint);
}
device -= rpc_count;
#endif
if (device < (int)model.devices.size()) {
return ggml_backend_dev_buffer_type(model.devices[device]);
}
device -= (int)model.devices.size();
#if defined(GGML_USE_METAL)
buft = ggml_backend_metal_buffer_type();
#elif defined(GGML_USE_VULKAN)
#if defined(GGML_USE_VULKAN)
buft = ggml_backend_vk_buffer_type(device);
#elif defined(GGML_USE_SYCL)
buft = ggml_backend_sycl_buffer_type(device);
@ -3556,18 +3523,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
}
static size_t llama_get_device_memory(const llama_model & model, int device) {
#if defined(GGML_USE_RPC)
int rpc_count = (int)model.rpc_servers.size();
if (device < rpc_count) {
size_t total;
size_t free;
const char * endpoint = model.rpc_servers[device].c_str();
ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
return free;
}
device = device - rpc_count;
#endif
if (device < (int)model.devices.size()) {
ggml_backend_dev_t dev = model.devices[device];
size_t total;
@ -8970,48 +8925,40 @@ static bool llm_load_tensors(
llama_buf_map bufs;
bufs.reserve(n_max_backend_buffer);
// only the mmap region containing the tensors in the model is mapped to the backend buffer
// this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
// this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(model, true)) {
// check if this backend device supports buffer_from_host_ptr
// when using a host buffer as the CPU bakcend buffer, use the CPU device to prioritize using buffer_from_host_ptr over the host buffer
ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft == llama_default_buffer_type_cpu(model, true) ? ggml_backend_cpu_buffer_type() : buft);
bool buffer_from_host_ptr_supported = false;
if (dev) {
ggml_backend_dev_props props;
ggml_backend_dev_get_props(dev, &props);
buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
}
if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported) {
for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
// only the mmap region containing the tensors in the model is mapped to the backend buffer
// this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
// this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
void * addr = nullptr;
size_t first, last;
size_t first, last; // NOLINT
ml.get_mapping_range(&first, &last, &addr, idx, ctx);
if (first >= last) {
continue;
}
ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
if (buf == nullptr) {
throw std::runtime_error("unable to allocate backend CPU buffer");
}
model.bufs.push_back(buf);
bufs.emplace(idx, buf);
}
}
#ifdef GGML_USE_METAL
else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
const size_t max_size = ggml_get_max_tensor_size(ctx);
void * addr = nullptr;
size_t first, last;
ml.get_mapping_range(&first, &last, &addr, idx, ctx);
if (first >= last) {
continue;
}
ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size);
ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
if (buf == nullptr) {
throw std::runtime_error("unable to allocate backend metal buffer");
throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
}
model.bufs.push_back(buf);
bufs.emplace(idx, buf);
}
}
#endif
else {
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
if (buf == nullptr) {
throw std::runtime_error("unable to allocate backend buffer");
throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
}
model.bufs.push_back(buf);
if (use_mlock && ggml_backend_buffer_is_host(buf)) {
@ -17138,17 +17085,19 @@ static void llama_graph_compute(
int n_threads,
ggml_threadpool * threadpool) {
if (lctx.backend_cpu != nullptr) {
ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
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);
}
#ifdef GGML_USE_BLAS
if (lctx.backend_blas != nullptr) {
ggml_backend_blas_set_n_threads(lctx.backend_blas, n_threads);
}
#endif
ggml_backend_sched_graph_compute_async(lctx.sched, gf);
// set the number of threads for all the backends
for (const auto & set_n_threads_fn : lctx.set_n_threads_fns) {
set_n_threads_fn.second(set_n_threads_fn.first, n_threads);
}
auto err = ggml_backend_sched_graph_compute_async(lctx.sched, gf);
if (err != GGML_STATUS_SUCCESS) {
LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, err);
}
// fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
}
@ -17964,10 +17913,9 @@ static void llama_tensor_dequantize_internal(
}
float * f32_output = (float *) output.data();
ggml_type_traits_t qtype;
const ggml_type_traits * qtype = ggml_get_type_traits(tensor->type);
if (ggml_is_quantized(tensor->type)) {
qtype = ggml_internal_get_type_traits(tensor->type);
if (qtype.to_float == NULL) {
if (qtype->to_float == NULL) {
throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
}
} else if (tensor->type != GGML_TYPE_F16 &&
@ -17981,7 +17929,7 @@ static void llama_tensor_dequantize_internal(
} else if (tensor->type == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
} else if (ggml_is_quantized(tensor->type)) {
qtype.to_float(tensor->data, f32_output, nelements);
qtype->to_float(tensor->data, f32_output, nelements);
} else {
GGML_ABORT("fatal error"); // unreachable
}
@ -18017,7 +17965,7 @@ static void llama_tensor_dequantize_internal(
} else if (typ == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
} else {
qtype.to_float(inbuf, outbuf, nels);
qtype->to_float(inbuf, outbuf, nels);
}
};
workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
@ -19111,16 +19059,21 @@ bool llama_supports_mlock(void) {
}
bool llama_supports_gpu_offload(void) {
#if defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
#if defined(GGML_USE_CLBLAST) || defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
return true;
#else
return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr ||
ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr;
ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr ||
llama_supports_rpc();
#endif
}
bool llama_supports_rpc(void) {
return ggml_backend_reg_by_name("RPC") != nullptr;
}
void llama_backend_init(void) {
ggml_time_init();
@ -19195,14 +19148,51 @@ struct llama_model * llama_load_model_from_file(
model->rpc_servers.push_back(servers);
}
// add RPC devices
if (!model->rpc_servers.empty()) {
ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
if (!rpc_reg) {
LLAMA_LOG_ERROR("%s: failed to find RPC backend\n", __func__);
llama_free_model(model);
return nullptr;
}
// ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
using ggml_backend_rpc_add_device_t = ggml_backend_dev_t (*)(const char *);
ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
if (!ggml_backend_rpc_add_device_fn) {
LLAMA_LOG_ERROR("%s: failed to find RPC device add function\n", __func__);
llama_free_model(model);
return nullptr;
}
for (const std::string & server : model->rpc_servers) {
ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
if (dev) {
model->devices.push_back(dev);
} else {
LLAMA_LOG_ERROR("%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
llama_free_model(model);
return nullptr;
}
}
}
// 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);
// skip the CPU backend since it is handled separately
if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU_FULL) {
model->devices.push_back(dev);
switch (ggml_backend_dev_type(dev)) {
case GGML_BACKEND_DEVICE_TYPE_CPU:
case GGML_BACKEND_DEVICE_TYPE_CPU_FULL:
// skip CPU backends since they are `handled separately
break;
case GGML_BACKEND_DEVICE_TYPE_GPU:
case GGML_BACKEND_DEVICE_TYPE_GPU_FULL:
model->devices.push_back(dev);
break;
}
}
@ -19214,7 +19204,7 @@ struct llama_model * llama_load_model_from_file(
} else if (status == -2) {
LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
}
delete model;
llama_free_model(model);
return nullptr;
}
@ -19397,34 +19387,7 @@ struct llama_context * llama_new_context_with_model(
main_gpu -= (int)model->devices.size();
}
#if defined(GGML_USE_RPC)
if (model->n_gpu_layers > 0) {
for (const auto & endpoint : model->rpc_servers) {
ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
}
if (main_gpu >= (int)model->rpc_servers.size()) {
main_gpu -= (int)model->rpc_servers.size();
}
#endif
#if defined(GGML_USE_METAL)
if (model->n_gpu_layers > 0) {
ctx->backend_metal = ggml_backend_metal_init();
if (ctx->backend_metal == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(ctx->backend_metal);
}
#elif defined(GGML_USE_VULKAN)
#if defined(GGML_USE_VULKAN)
if (model->split_mode == LLAMA_SPLIT_MODE_ROW) {
LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__);
llama_free(ctx);
@ -19507,14 +19470,19 @@ struct llama_context * llama_new_context_with_model(
}
#endif
#ifdef GGML_USE_BLAS
ctx->backend_blas = ggml_backend_blas_init();
if (ctx->backend_blas == nullptr) {
LLAMA_LOG_WARN("%s: failed to initialize BLAS backend\n", __func__);
} else {
ctx->backends.push_back(ctx->backend_blas);
// add other backends (such as BLAS)
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
}
#endif
ctx->backend_cpu = ggml_backend_cpu_init();
if (ctx->backend_cpu == nullptr) {
@ -19524,6 +19492,18 @@ struct llama_context * llama_new_context_with_model(
}
ctx->backends.push_back(ctx->backend_cpu);
// create a list of the set_n_threads functions in the backends
for (auto * backend : ctx->backends) {
ggml_backend_dev_t dev = ggml_backend_get_device(backend);
ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
if (reg) {
auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
if (ggml_backend_set_n_threads_fn) {
ctx->set_n_threads_fns.emplace_back(backend, ggml_backend_set_n_threads_fn);
}
}
}
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);