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

ggml : add metal backend registry / device (#9713)

Browse source 
* ggml : add metal backend registry / device

ggml-ci

* metal : fix names [no ci]

* metal : global registry and device instances

ggml-ci

* cont : alternative initialization of global objects

ggml-ci

* llama : adapt to backend changes

ggml-ci

* fixes

* metal : fix indent

* metal : fix build when MTLGPUFamilyApple3 is not available

ggml-ci

* fix merge

* metal : avoid unnecessary singleton accesses

ggml-ci

* metal : minor fix [no ci]

* metal : g_state -> g_ggml_ctx_dev_main [no ci]

* metal : avoid reference of device context in the backend context

ggml-ci

* metal : minor [no ci]

* metal : fix maxTransferRate check

* metal : remove transfer rate stuff

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov 2024-10-07 18:27:51 +03:00 • committed by GitHub
parent 96b6912103
commit d5ac8cf2f2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 535 additions and 284 deletions
Download patch file Download diff file Expand all files Collapse all files

72
src/llama.cpp
View file

@ -26,10 +26,6 @@
# include "ggml-blas.h"
#endif
#ifdef GGML_USE_METAL
# include "ggml-metal.h"
#endif
// TODO: replace with ggml API call
#define QK_K 256
@ -3292,9 +3288,6 @@ 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
@ -3420,9 +3413,7 @@ static int llama_get_device_count(const llama_model & model) {
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();
@ -3488,9 +3479,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_
}
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);
@ -8918,48 +8907,39 @@ 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
ggml_backend_dev_t dev = ggml_backend_buft_get_device(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)) {
@ -19041,7 +19021,7 @@ bool llama_supports_mlock(void) {
}
bool llama_supports_gpu_offload(void) {
#if defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
#if defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
return true;
@ -19344,17 +19324,7 @@ struct llama_context * llama_new_context_with_model(
}
#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);