Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-15 23:00:46 +01:00
Code Issues Packages Projects Releases Wiki Activity

SYCL: fix assertions and add proper comments

Browse Source
This commit is contained in:
Akarshan Biswas 2024-12-15 19:11:43 +05:30
parent da40c42062
commit f8603b0cc0
No known key found for this signature in database
GPG Key ID: 52A578A14B32134D
2 changed files with 17 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-sycl/common.cpp
View File

@ -11,6 +11,8 @@
// //
#include "common.hpp" #include "common.hpp"
#include "ggml-backend-impl.h"
#include "ggml-impl.h" #include "ggml-impl.h"
int get_current_device_id() { int get_current_device_id() {

19
ggml/src/ggml-sycl/ggml-sycl.cpp
View File

@ -2352,16 +2352,27 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
// GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr); // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
} else if (ggml_backend_buffer_is_sycl(src->buffer) || ggml_backend_buffer_is_sycl_split(src->buffer)) { } else if (ggml_backend_buffer_is_sycl(src->buffer) || ggml_backend_buffer_is_sycl_split(src->buffer)) {
if (!ggml_backend_buffer_is_sycl_split(src->buffer)){ if (!ggml_backend_buffer_is_sycl_split(src->buffer)){
// Tensor is already on the device, what kind to choose here? // If buffer is not a SYCL split buffer
/*
What memcpy_direction kind we need here?
Refer: dpct/helper.hpp:
enum memcpy_direction
{
host_to_host,
host_to_device,
device_to_host,
device_to_device,
automatic
};
*/
kind = dpct::device_to_device; kind = dpct::device_to_device;
src_ptr = (char *) src->data; src_ptr = (char *) src->data;
} }
else { else {
/* /*
This assertion seems to me that split buffers aren't supported in SYCL If buffer is a SYCL split buffer
Use ggml_abort()?
*/ */
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src->buffer) || (i1_low == 0 && i1_high == src->ne[1])); GGML_ASSERT(i1_low == 0 && i1_high == src->ne[1]);
kind = dpct::device_to_device; kind = dpct::device_to_device;
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra; ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
int id; int id;