root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

vulkan : add backend registry / device interfaces (#9721)
Some checks are pending
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-cuda.Dockerfile platforms:linux/amd64 tag:full-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-musa.Dockerfile platforms:linux/amd64 tag:full-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full.Dockerfile platforms:linux/amd64,linux/arm64 tag:full]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-cuda.Dockerfile platforms:linux/amd64 tag:light-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-intel.Dockerfile platforms:linux/amd64 tag:light-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-musa.Dockerfile platforms:linux/amd64 tag:light-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli.Dockerfile platforms:linux/amd64,linux/arm64 tag:light]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-cuda.Dockerfile platforms:linux/amd64 tag:server-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-intel.Dockerfile platforms:linux/amd64 tag:server-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-musa.Dockerfile platforms:linux/amd64 tag:server-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server.Dockerfile platforms:linux/amd64,linux/arm64 tag:server]) (push) Waiting to run
Details
Nix CI / nix-eval (macos-latest) (push) Waiting to run
Details
Nix CI / nix-eval (ubuntu-latest) (push) Waiting to run
Details
Nix CI / nix-build (macos-latest) (push) Waiting to run
Details
Nix CI / nix-build (ubuntu-latest) (push) Waiting to run
Details
flake8 Lint / Lint (push) Waiting to run
Details

Browse Source 
* vulkan : add backend registry / device interfaces

* llama : print devices used on model load
This commit is contained in:
Diego Devesa 2024-10-17 02:46:58 +02:00 committed by GitHub
parent 2194200278
commit f010b77a37
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 226 additions and 123 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/include/ggml-vulkan.h
View File

@ -24,6 +24,8 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
#ifdef __cplusplus
}
#endif

9
ggml/src/ggml-backend.cpp
View File

@ -538,6 +538,10 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
@ -557,6 +561,9 @@ struct ggml_backend_registry {
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
@ -564,7 +571,7 @@ struct ggml_backend_registry {
register_backend(ggml_backend_rpc_reg());
#endif
// TODO: sycl, vulkan, kompute, cann
// TODO: sycl, kompute, cann
register_backend(ggml_backend_cpu_reg());
}

284
ggml/src/ggml-vulkan.cpp
View File

@ -1941,7 +1941,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
if (device->fp16) {
device_extensions.push_back("VK_KHR_shader_float16_int8");
}
device->name = device->properties.deviceName.data();
device->name = GGML_VK_NAME + std::to_string(idx);
device_create_info = {
vk::DeviceCreateFlags(),
@ -1968,7 +1968,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->buffer_type = {
/* .iface = */ ggml_backend_vk_buffer_type_interface,
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), idx),
/* .context = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
};
@ -6378,7 +6378,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), 0),
/* .context = */ nullptr,
};
@ -6581,9 +6581,135 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
UNUSED(backend);
}
static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
// ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
// TODO: enable async and synchronize
static ggml_backend_i ggml_backend_vk_interface = {
/* .get_name = */ ggml_backend_vk_name,
/* .free = */ ggml_backend_vk_free,
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_vk_graph_compute,
/* .supports_op = */ NULL,
/* .supports_buft = */ NULL,
/* .offload_op = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_vk_guid() {
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
return &guid;
}
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
ggml_vk_init(ctx, dev_num);
ggml_backend_t vk_backend = new ggml_backend {
/* .guid = */ ggml_backend_vk_guid(),
/* .interface = */ ggml_backend_vk_interface,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
/* .context = */ ctx,
};
return vk_backend;
}
bool ggml_backend_is_vk(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
}
int ggml_backend_vk_get_device_count() {
return ggml_vk_get_device_count();
}
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
int dev_idx = vk_instance.device_indices[device];
ggml_vk_get_device_description(dev_idx, description, description_size);
}
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
*total = heap.size;
*free = heap.size;
break;
}
}
}
//////////////////////////
struct ggml_backend_vk_device_context {
int device;
std::string name;
std::string description;
};
static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ctx->name.c_str();
}
static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_vk_device_get_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)device->context;
ggml_backend_vk_get_device_memory(ctx->device, free, total);
}
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ggml_backend_vk_buffer_type(ctx->device);
}
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(ggml_backend_dev_t dev) {
UNUSED(dev);
return ggml_backend_vk_host_buffer_type();
}
static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
UNUSED(dev);
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
}
static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_vk_device_get_name(dev);
props->description = ggml_backend_vk_device_get_description(dev);
props->type = ggml_backend_vk_device_get_type(dev);
ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* async */ false,
/* host_buffer */ true,
/* events */ false,
};
}
static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
UNUSED(params);
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ggml_backend_vk_init(ctx->device);
}
static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
@ -6701,97 +6827,101 @@ static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tenso
return false;
}
UNUSED(backend);
UNUSED(dev);
}
static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
static bool ggml_backend_vk_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
return false;
}
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
return buft_ctx->device->idx == ctx->device;
}
static bool ggml_backend_vk_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
const int min_batch_size = 32;
return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
(op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
UNUSED(backend);
UNUSED(dev);
}
static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
return false;
}
ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
return buft_ctx->device == ctx->device;
}
// TODO: enable async and synchronize
static ggml_backend_i ggml_backend_vk_interface = {
/* .get_name = */ ggml_backend_vk_name,
/* .free = */ ggml_backend_vk_free,
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_vk_graph_compute,
/* .supports_op = */ ggml_backend_vk_supports_op,
/* .supports_buft = */ ggml_backend_vk_supports_buft,
/* .offload_op = */ ggml_backend_vk_offload_op,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
static const struct ggml_backend_device_i ggml_backend_vk_device_i = {
/* .get_name = */ ggml_backend_vk_device_get_name,
/* .get_description = */ ggml_backend_vk_device_get_description,
/* .get_memory = */ ggml_backend_vk_device_get_memory,
/* .get_type = */ ggml_backend_vk_device_get_type,
/* .get_props = */ ggml_backend_vk_device_get_props,
/* .init_backend = */ ggml_backend_vk_device_init,
/* .get_buffer_type = */ ggml_backend_vk_device_get_buffer_type,
/* .get_host_buffer_type = */ ggml_backend_vk_device_get_host_buffer_type,
/* .buffer_from_host_ptr = */ NULL,
/* .supports_op = */ ggml_backend_vk_device_supports_op,
/* .supports_buft = */ ggml_backend_vk_device_supports_buft,
/* .offload_op = */ ggml_backend_vk_device_offload_op,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_vk_guid() {
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
return &guid;
static const char * ggml_backend_vk_reg_get_name(ggml_backend_reg_t reg) {
UNUSED(reg);
return GGML_VK_NAME;
}
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
static size_t ggml_backend_vk_reg_get_device_count(ggml_backend_reg_t reg) {
UNUSED(reg);
return ggml_backend_vk_get_device_count();
}
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
ggml_vk_init(ctx, dev_num);
static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg, size_t device) {
static std::vector<ggml_backend_dev_t> devices;
ggml_backend_t vk_backend = new ggml_backend {
/* .guid = */ ggml_backend_vk_guid(),
/* .interface = */ ggml_backend_vk_interface,
/* .device = */ nullptr,
static bool initialized = false;
{
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
for (size_t i = 0; i < ggml_backend_vk_get_device_count(); i++) {
ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
char desc[256];
ggml_backend_vk_get_device_description(i, desc, sizeof(desc));
ctx->device = i;
ctx->name = GGML_VK_NAME + std::to_string(i);
ctx->description = desc;
devices.push_back(new ggml_backend_device {
/* .iface = */ ggml_backend_vk_device_i,
/* .reg = */ reg,
/* .context = */ ctx,
});
}
initialized = true;
}
}
GGML_ASSERT(device < devices.size());
return devices[device];
}
static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
/* .get_name = */ ggml_backend_vk_reg_get_name,
/* .get_device_count = */ ggml_backend_vk_reg_get_device_count,
/* .get_device = */ ggml_backend_vk_reg_get_device,
/* .get_proc_address = */ NULL,
};
return vk_backend;
}
ggml_backend_reg_t ggml_backend_vk_reg() {
static ggml_backend_reg reg = {
/* .iface = */ ggml_backend_vk_reg_i,
/* .context = */ nullptr,
};
bool ggml_backend_is_vk(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
}
int ggml_backend_vk_get_device_count() {
return ggml_vk_get_device_count();
}
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
ggml_vk_get_device_description(device, description, description_size);
}
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
*total = heap.size;
*free = heap.size;
break;
}
}
return &reg;
}
// Extension availability

54
src/llama.cpp
View File

@ -8,9 +8,7 @@
#include "ggml-alloc.h"
#include "ggml-backend.h"
#if defined(GGML_USE_VULKAN)
# include "ggml-vulkan.h"
#elif defined(GGML_USE_SYCL)
#if defined(GGML_USE_SYCL)
# include "ggml-sycl.h"
#elif defined(GGML_USE_KOMPUTE)
# include "ggml-kompute.h"
@ -3418,8 +3416,6 @@ static int llama_get_device_count(const llama_model & model) {
#if defined(GGML_USE_SYCL)
count += ggml_backend_sycl_get_device_count();
#elif defined(GGML_USE_VULKAN)
count += ggml_backend_vk_get_device_count();
#elif defined(GGML_USE_CANN)
count += ggml_backend_cann_get_device_count();
#endif
@ -3451,10 +3447,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
}
#elif defined(GGML_USE_CPU_HBM)
buft = ggml_backend_cpu_hbm_buffer_type();
#elif defined(GGML_USE_VULKAN)
if (host_buffer) {
buft = ggml_backend_vk_host_buffer_type();
}
#endif
if (buft == nullptr) {
@ -3473,9 +3465,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_
}
device -= (int)model.devices.size();
#if defined(GGML_USE_VULKAN)
buft = ggml_backend_vk_buffer_type(device);
#elif defined(GGML_USE_SYCL)
#if defined(GGML_USE_SYCL)
buft = ggml_backend_sycl_buffer_type(device);
#elif defined(GGML_USE_KOMPUTE)
buft = ggml_backend_kompute_buffer_type(device);
@ -3535,11 +3525,6 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
size_t free;
ggml_backend_sycl_get_device_memory(device, &free, &total);
return free;
#elif defined(GGML_USE_VULKAN)
size_t total;
size_t free;
ggml_backend_vk_get_device_memory(device, &free, &total);
return free;
#elif defined(GGML_USE_CANN)
size_t total;
size_t free;
@ -19095,8 +19080,7 @@ bool llama_supports_mlock(void) {
}
bool llama_supports_gpu_offload(void) {
#if defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
#if 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
@ -19227,10 +19211,15 @@ struct llama_model * llama_load_model_from_file(
case GGML_BACKEND_DEVICE_TYPE_GPU:
case GGML_BACKEND_DEVICE_TYPE_GPU_FULL:
{
size_t free, total; // NOLINT
ggml_backend_dev_memory(dev, &free, &total);
LLAMA_LOG_INFO("%s: using device %s (%s) - %zu MiB free\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), free/1024/1024);
model->devices.push_back(dev);
break;
}
}
}
int status = llama_model_load(path_model, *model, params);
GGML_ASSERT(status <= 0);
@ -19423,32 +19412,7 @@ struct llama_context * llama_new_context_with_model(
main_gpu -= (int)model->devices.size();
}
#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);
return nullptr;
}
if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
ggml_backend_t backend = ggml_backend_vk_init(main_gpu);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
} else {
for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
ggml_backend_t backend = ggml_backend_vk_init(device);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Vulkan%d backend\n", __func__, device);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
}
#elif defined(GGML_USE_SYCL)
#if defined(GGML_USE_SYCL)
// with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
ggml_backend_t backend = ggml_backend_sycl_init(main_gpu);