ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 15:14:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Clearing padding

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-05-20 16:59:31 +03:00
parent 651fed0848
commit 5252e95552
3 changed files with 12 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/include/ggml-backend.h
View File

@@ -22,7 +22,7 @@ extern "C" {
GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size);
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft);
GGML_API GGML_CALL size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
GGML_API GGML_CALL size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
// buffer
@@ -39,7 +39,7 @@ extern "C" {
GGML_API GGML_CALL void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);

4
ggml/src/ggml-backend.c
View File

@@ -35,7 +35,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
return SIZE_MAX;
}
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
// get_alloc_size is optional, defaults to ggml_nbytes
if (buft->iface.get_alloc_size) {
size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -114,7 +114,7 @@ size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
}
size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor) {
return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
}

10
ggml/src/ggml-cuda.cu
View File

@@ -2101,13 +2101,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
// If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q.
// But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data.
// Therefore, in such cases use cuBLAS.
const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
&& ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
bool use_mul_mat_q = ggml_is_quantized(src0->type)
bool use_mul_mat_q = ggml_is_quantized(src0->type) && !bad_padding_clear
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
// if mmvq is available it's a better choice than dmmv: