vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-04-28 03:30:20 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

fixed smartcache bug when used with images

Browse source
This commit is contained in:
Concedo 2026-01-02 00:35:05 +08:00
parent 774841ffd6
commit bfa2ae7744
3 changed files with 14 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

2
embd_res/klite.embd
View file

@ -4442,7 +4442,7 @@ Current version indicated by LITEVER below.
}
if (typeof obj === 'object') {
const result = {};
for (const key in obj) {
for (let key in obj) {
if (obj.hasOwnProperty(key)) {
result[key] = replaceStringsInObject(obj[key], src, tgt);
}

13
gpttype_adapter.cpp
View file

@ -4019,6 +4019,10 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
for(int i=0;i<savestate_limit;++i)
{
bool target_usable = FullyContainedPrefix(savestates[i].savestate_context_tokens,embd_inp);
if(savestates[i].media_signature!=media_composite_image_signature)
{
target_usable = false;
}
int target_len = savestates[i].savestate_context_tokens.size();
if(target_usable && target_len>bestlen)
{
@ -4085,6 +4089,10 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
for(int i=0;i<savestate_limit;++i)
{
float similaritybeat = ComputePrefixMatchPercent(savestates[i].savestate_context_tokens,embd_inp);
if(savestates[i].media_signature!=media_composite_image_signature)
{
continue;
}
if(similaritybeat > similarity_threshold || (shiftable && CanContextShift(savestates[i].savestate_context_tokens, embd_inp, inputs.max_length, nctx)))
{
//found a match. save to the oldest slot thats not the one we are loading
@ -5110,6 +5118,7 @@ size_t gpttype_save_state_kv(int slot)
savestates[slot].savestate_context_tokens.clear();
savestates[slot].current_savestate_size = 0;
savestates[slot].current_draft_savestate_size = 0;
savestates[slot].media_signature = "";
}
size_t newsize = llama_state_get_size(llama_ctx_v4);
try {
@ -5127,6 +5136,7 @@ size_t gpttype_save_state_kv(int slot)
totalbytes += res;
savestates[slot].current_savestate_size = newsize;
savestates[slot].savestate_context_tokens = current_context_tokens;
savestates[slot].media_signature = media_composite_image_signature;
int maxedpos = llama_memory_seq_pos_max(llama_get_memory(llama_ctx_v4),0);
if(maxedpos > 0 && savestates[slot].savestate_context_tokens.size() > maxedpos && savestates[slot].savestate_context_tokens.size()-maxedpos<=2)
{
@ -5211,6 +5221,7 @@ bool gpttype_clear_state_kv(bool shrink)
}
savestates[slot].savestate_context_tokens.clear();
savestates[slot].current_savestate_size = 0;
savestates[slot].media_signature = "";
if(draft_ctx && savestates[slot].current_draft_savestate_size>0)
{
savestates[slot].current_draft_savestate_buffer.clear();
@ -5240,7 +5251,7 @@ int get_identical_existing_slot() //returns slot number of slot containing exact
int currctxsize = current_context_tokens.size();
for(int i=0;i<savestate_limit;++i)
{
if(savestates[i].savestate_context_tokens.size() == currctxsize)
if(savestates[i].savestate_context_tokens.size() == currctxsize && savestates[i].media_signature==media_composite_image_signature)
{
bool is_identical = true;
const auto& slot_tokens = savestates[i].savestate_context_tokens;

1
otherarch/otherarch.h
View file

@ -538,6 +538,7 @@ struct savestate_data
std::vector<uint8_t> current_draft_savestate_buffer;
std::vector<gpt_vocab::id> savestate_context_tokens; //for context clones
int64_t last_used = 0; //unix timestamp, updated on save or load
std::string media_signature = "";
};
const float default_norm_eps = 1e-5f;