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

Final touches

Browse source
This commit is contained in:
Akarshan Biswas 2025-02-03 21:08:25 +05:30
parent 52b0652601
commit 0b602f0ecd
No known key found for this signature in database
GPG key ID: 52A578A14B32134D
2 changed files with 12 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

1
ggml/src/ggml-sycl/binbcast.cpp
View file

@ -1,5 +1,4 @@
#include "binbcast.hpp" #include "binbcast.hpp"
#include "common.hpp"
static __dpct_inline__ float op_repeat(const float a, const float b) { static __dpct_inline__ float op_repeat(const float a, const float b) {
return b; return b;

13
ggml/src/ggml-sycl/ggml-sycl.cpp
View file

@ -309,6 +309,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor, ggml_tensor *tensor,
const void *data, size_t offset, const void *data, size_t offset,
size_t size) try { size_t size) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@ -333,6 +334,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *tensor, const ggml_tensor *tensor,
void *data, size_t offset, void *data, size_t offset,
size_t size) try { size_t size) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@ -361,6 +363,7 @@ static bool
ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer, ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *src, const ggml_tensor *src,
ggml_tensor *dst) try { ggml_tensor *dst) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
if (ggml_backend_buffer_is_sycl(src->buffer)) { if (ggml_backend_buffer_is_sycl(src->buffer)) {
ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context; ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
ggml_backend_sycl_buffer_context * dst_ctx = (ggml_backend_sycl_buffer_context *)dst->buffer->context; ggml_backend_sycl_buffer_context * dst_ctx = (ggml_backend_sycl_buffer_context *)dst->buffer->context;
@ -418,7 +421,8 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer, static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
uint8_t value) try { uint8_t value) try {
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
ggml_sycl_set_device(ctx->device); ggml_sycl_set_device(ctx->device);
queue_ptr stream = ctx->stream; queue_ptr stream = ctx->stream;
@ -465,6 +469,7 @@ static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_t
static ggml_backend_buffer_t static ggml_backend_buffer_t
ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
size_t size) try { size_t size) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
ggml_sycl_set_device(buft_ctx->device); ggml_sycl_set_device(buft_ctx->device);
const queue_ptr stream = buft_ctx->stream; const queue_ptr stream = buft_ctx->stream;
@ -708,6 +713,7 @@ static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buff
static void static void
ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor) try { ggml_tensor *tensor) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context; ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
@ -791,6 +797,7 @@ static void
ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor, const void *data, ggml_tensor *tensor, const void *data,
size_t offset, size_t size) try { size_t offset, size_t size) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
// split tensors must always be set in their entirety at once // split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0); GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor)); GGML_ASSERT(size == ggml_nbytes(tensor));
@ -844,6 +851,7 @@ static void
ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer, ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *tensor, void *data, const ggml_tensor *tensor, void *data,
size_t offset, size_t size) try { size_t offset, size_t size) try {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
// split tensors must always be set in their entirety at once // split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0); GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor)); GGML_ASSERT(size == ggml_nbytes(tensor));
@ -894,6 +902,7 @@ catch (sycl::exception const &exc) {
} }
static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
GGML_UNUSED(buffer); GGML_UNUSED(buffer);
GGML_UNUSED(value); GGML_UNUSED(value);
} }
@ -1017,10 +1026,12 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
} }
static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
ggml_sycl_host_free(buffer->context); ggml_sycl_host_free(buffer->context);
} }
static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
void * ptr = ggml_sycl_host_malloc(size); void * ptr = ggml_sycl_host_malloc(size);
if (ptr == nullptr) { if (ptr == nullptr) {