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

Add back split buffer type checks

Browse source
This commit is contained in:
Akarshan Biswas 2025-02-01 12:06:58 +05:30
parent 7d8d689d39
commit 04d8b038b8
No known key found for this signature in database
GPG key ID: 52A578A14B32134D
6 changed files with 16 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

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

@ -52,6 +52,7 @@ static void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * d
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_I32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);

2
ggml/src/ggml-sycl/cpy.hpp
View file

@ -8,4 +8,4 @@ typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1);
#endif // GGML_SYCL_CPY_HPP
#endif // GGML_SYCL_CPY_HPP

8
ggml/src/ggml-sycl/ggml-sycl.cpp
View file

@ -1920,6 +1920,8 @@ static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor
GGML_ASSERT(dst->src[0]->nb[0] == ggml_type_size(dst->src[0]->type));
GGML_ASSERT(dst->src[1]->nb[0] == ggml_type_size(dst->src[1]->type));
GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->src[1]->buffer));
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
switch (dst->src[0]->type) {
case GGML_TYPE_F16:
@ -2087,6 +2089,7 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor *d
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
const int32_t * opts = (const int32_t *)dst->op_params;
enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
@ -2125,6 +2128,7 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor *d
inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
const int64_t ne = ggml_nelements(dst->src[0]);
dpct::queue_ptr main_stream = ctx.stream();
@ -2138,6 +2142,7 @@ inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor *
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
@ -2152,6 +2157,7 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_ten
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t ne01 = dst->src[0]->ne[1];
@ -2169,6 +2175,7 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * ds
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
@ -2189,6 +2196,7 @@ inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor *dst
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
float min;
float max;

1
ggml/src/ggml-sycl/im2col.cpp
View file

@ -86,6 +86,7 @@ void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F16);
GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->src[1]->buffer->buft->iface.get_name(dst->src[1]->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
const int32_t s0 = ((const int32_t*)(dst->op_params))[0];

3
ggml/src/ggml-sycl/norm.cpp
View file

@ -315,6 +315,7 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
@ -333,6 +334,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
int num_groups = dst->op_params[0];
@ -350,6 +352,7 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);

2
ggml/src/ggml-sycl/rope.cpp
View file

@ -197,6 +197,8 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(dst->src[0]->type == dst->type);
GGML_ASSERT(strcmp(dst->src[1]->buffer->buft->iface.get_name(dst->src[1]->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t ne01 = dst->src[0]->ne[1];