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

add enough padding for alignment

Browse source
This commit is contained in:
FSSRepo 2023-11-24 12:48:32 -05:00
parent f4f0b06a9c
commit b13911f02c
1 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
examples/llava/clip.cpp
View file

@ -505,15 +505,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
const size_t offset = gguf_get_tensor_offset(ctx, i); const size_t offset = gguf_get_tensor_offset(ctx, i);
struct ggml_tensor * cur = ggml_get_tensor(meta, name); struct ggml_tensor * cur = ggml_get_tensor(meta, name);
size_t tensor_size = ggml_nbytes(cur); size_t tensor_size = ggml_nbytes(cur);
size_t padded_size = ggml_nbytes_pad(cur); buffer_size += tensor_size;
buffer_size += padded_size;
if (verbosity >= 3) { if (verbosity >= 3) {
printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i, printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i,
cur->n_dims, cur->name, tensor_size, padded_size, offset); cur->n_dims, cur->name, tensor_size, offset);
} }
} }
} }
buffer_size += n_tensors * 128 /* CLIP PADDING */;
clip_ctx * new_clip = new clip_ctx; clip_ctx * new_clip = new clip_ctx;
#ifdef CLIP_USE_CUBLAS #ifdef CLIP_USE_CUBLAS
new_clip->backend = ggml_backend_cuda_init(); new_clip->backend = ggml_backend_cuda_init();