vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

remove unused function

Browse source
This commit is contained in:
Concedo 2024-12-16 11:39:52 +08:00
parent 00d154b32b
commit bc297da91e
1 changed files with 0 additions and 64 deletions
Download patch file Download diff file Expand all files Collapse all files

64
gpttype_adapter.cpp
View file

@ -1804,70 +1804,6 @@ static bool kcpp_eval_image(llama_context * ctx_llama, float * img_embd, int num
}
return true;
}
static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, float * image_embd, int num_img_tokens,
int n_batch, int * n_past) {
auto image_size = clip_get_load_image_size(clp_ctx);
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
const int patch_size = 14 * 2;
const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0);
const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0);
auto img_tokens = num_img_tokens;
// llama_pos mrope_pos[img_tokens * 4];
std::vector<llama_pos> mrope_pos;
mrope_pos.resize(img_tokens * 4);
int st_pos_id = *n_past;
for (int y = 0; y < ph; y++)
{
for (int x = 0; x < pw; x++)
{
int i = y * pw + x;
mrope_pos[i] = st_pos_id;
mrope_pos[i + img_tokens] = st_pos_id + y;
mrope_pos[i + img_tokens * 2] = st_pos_id + x;
mrope_pos[i + img_tokens * 3] = 0;
}
}
st_pos_id += std::max(pw, ph);
int processed = 0;
std::vector<llama_pos> batch_mrope_pos;
batch_mrope_pos.resize(img_tokens * 4);
for (int i = 0; i < img_tokens; i += n_batch) {
int n_eval = img_tokens - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
// llama_pos batch_mrope_pos[n_eval * 4];
std::fill(batch_mrope_pos.begin(), batch_mrope_pos.end(), 0);
memcpy(batch_mrope_pos.data(), &mrope_pos[processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 1], &mrope_pos[img_tokens * 1 + processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos));
llama_batch batch = {
int32_t(n_eval), // n_tokens
nullptr, // token
(image_embd+i*n_embd), // embed
batch_mrope_pos.data(), // pos
nullptr, // n_seq_id
nullptr, // seq_id
nullptr, // logits
};
if (llama_decode(ctx_llama, batch)) {
fprintf(stderr, "\n%s : failed to eval image\n", __func__);
return false;
}
*n_past += n_eval;
processed += n_eval;
}
return true;
}
//given an old GGUF context and a new context that has some middle portion removed,
//find and remove the middle portion from the old context from the KV. Does not fast forward after this destructive action