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

[CANN] Add Ascend NPU backend (#6035)

Browse source 
* [CANN] Add Ascend NPU backend

Ascend is a full-stack AI computing infrastructure for industry
applications and services based on Huawei Ascend processors and
software.

CANN (Compute Architecture of Neural Networks), developped by
Huawei, is a heterogeneous computing architecture for AI.

Co-authored-by: wangshuai09 <391746016@qq.com>

* delete trailing whitespaces

* Modify the code based on review comment

* Rename LLAMA_CANN to GGML_CANN

* Make ggml-common.h private

* add ggml_cann prefix for acl funcs

* Add logging for CANN backend

* Delete Trailing whitespace

---------

Co-authored-by: wangshuai09 <391746016@qq.com>
This commit is contained in:
hipudding 2024-07-17 19:23:50 +08:00 • committed by GitHub
parent da3913d8f9
commit 1bdd8ae19f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
27 changed files with 10756 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

41
src/llama.cpp
View file

@ -19,6 +19,8 @@
# include "ggml-sycl.h"
#elif defined(GGML_USE_KOMPUTE)
# include "ggml-kompute.h"
#elif defined(GGML_USE_CANN)
# include "ggml-cann.h"
#endif
#ifdef GGML_USE_BLAS
@ -2079,6 +2081,8 @@ struct llama_state {
ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data);
#elif defined(GGML_USE_CUDA)
ggml_backend_cuda_log_set_callback(log_callback, log_callback_user_data);
#elif defined(GGML_USE_CANN)
ggml_backend_cann_log_set_callback(log_callback, log_callback_user_data);
#endif
}
@ -2889,6 +2893,8 @@ static size_t llama_get_device_count(const llama_model & model) {
count = ggml_backend_sycl_get_device_count();
#elif defined(GGML_USE_VULKAN)
count = ggml_backend_vk_get_device_count();
#elif defined(GGML_USE_CANN)
return ggml_backend_cann_get_device_count();
#endif
#if defined(GGML_USE_RPC)
count += model.rpc_servers.size();
@ -2921,6 +2927,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_
if (buft == nullptr) {
LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
}
#elif defined(GGML_USE_CANN)
buft = ggml_backend_cann_buffer_type(gpu);
#endif
if (buft == nullptr) {
@ -2981,6 +2989,11 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
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;
ggml_backend_cann_get_device_memory(device, &total, &free);
return free;
#else
return 1;
#endif
@ -18871,6 +18884,8 @@ size_t llama_max_devices(void) {
return GGML_SYCL_MAX_DEVICES;
#elif defined(GGML_USE_VULKAN)
return GGML_VK_MAX_DEVICES;
#elif defined(GGML_USE_CANN)
return GGML_CANN_MAX_DEVICES;
#else
return 1;
#endif
@ -19212,6 +19227,30 @@ struct llama_context * llama_new_context_with_model(
}
ctx->backends.push_back(backend);
}
#elif defined(GGML_USE_CANN)
// with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
// TODO: ggml_backend_cann is not support split tensor now, just leave code here.
if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
ggml_backend_t backend = ggml_backend_cann_init(model->main_gpu);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, model->main_gpu);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
} else {
// LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
// TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version.
for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) {
ggml_backend_t backend = ggml_backend_cann_init(device);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
}
#endif
#ifdef GGML_USE_BLAS
@ -21789,6 +21828,8 @@ void llama_log_set(ggml_log_callback log_callback, void * user_data) {
ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
#elif defined(GGML_USE_CUDA)
ggml_backend_cuda_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
#elif defined(GGML_USE_CANN)
ggml_backend_cann_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
#endif
}