Skip to content

cuda : fix tensor size calculation for non-split buffer #5145

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jan 26, 2024
Merged

cuda : fix tensor size calculation for non-split buffer #5145

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion ggml-backend.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,9 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
// get_alloc_size is optional, defaults to ggml_nbytes
if (buft->iface.get_alloc_size) {
return buft->iface.get_alloc_size(buft, tensor);
size_t size = buft->iface.get_alloc_size(buft, tensor);
assert(size >= ggml_nbytes(tensor));
return size;
}
return ggml_nbytes(tensor);
}
Expand Down
19 changes: 5 additions & 14 deletions ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9790,8 +9790,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
// TODO: mmq/mmv support
#endif

const int64_t nb11 = src1->nb[1];
const int64_t nb1 = dst->nb[1];
const size_t nb11 = src1->nb[1];
const size_t nb1 = dst->nb[1];

const struct ggml_tensor * ids = src0;
const int32_t id = ((int32_t *) dst->op_params)[0];
Expand Down Expand Up @@ -10304,15 +10304,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t

if (ggml_is_quantized(tensor->type)) {
// initialize padding to 0 to avoid possible NaN values
int64_t row_low = 0;
int64_t row_high = ggml_nrows(tensor);
int64_t nrows_split = row_high - row_low;

size_t original_size = ggml_nbytes_split(tensor, nrows_split);
size_t original_size = ggml_nbytes(tensor);
size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);

if (padded_size > original_size && tensor->view_src == nullptr) {
CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
}
}
}
Expand Down Expand Up @@ -10415,12 +10411,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend
}

GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
int64_t row_low = 0;
int64_t row_high = ggml_nrows(tensor);
int64_t nrows_split = row_high - row_low;

size_t size = ggml_nbytes_split(tensor, nrows_split);

size_t size = ggml_nbytes(tensor);
int64_t ne0 = tensor->ne[0];

if (ggml_is_quantized(tensor->type)) {
Expand Down