Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

smartcache bug seems to be fixed

Browse Source
This commit is contained in:
Concedo
2026-02-28 18:08:54 +08:00
parent 4e358265a3
commit 0b76f73fc2
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpttype_adapter.cpp
+6 -3
View File
@@ -5197,10 +5197,12 @@ size_t gpttype_save_state_kv(int slot)
savestates[slot].savestate_context_tokens = current_context_tokens; savestates[slot].savestate_context_tokens = current_context_tokens;
savestates[slot].media_signature = media_composite_image_signature; savestates[slot].media_signature = media_composite_image_signature;
int maxedpos = llama_memory_seq_pos_max(llama_get_memory(llama_ctx_v4),0); 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) //kcpp: so maxedpos appears to always be equal to ctx tokens - 2, if savestate_ctx_tokens > maxedpos + 2 then trim excess
if(maxedpos > 0 && savestates[slot].savestate_context_tokens.size() > maxedpos + 2)
{ {
//dirty hack for the memory actually being off by 1 or 2, correct the state //dirty hack for the memory actually being off, correct the state
while(savestates[slot].savestate_context_tokens.size() > maxedpos) printf("\nSaveState inconsistency fix, trimming from %d to %d\n",savestates[slot].savestate_context_tokens.size(),maxedpos+2);
while(savestates[slot].savestate_context_tokens.size() > maxedpos+2)
{ {
savestates[slot].savestate_context_tokens.pop_back(); savestates[slot].savestate_context_tokens.pop_back();
} }
@@ -5244,6 +5246,7 @@ bool gpttype_load_state_kv(int slot)
if (savestates[slot].current_savestate_buffer.empty()) { if (savestates[slot].current_savestate_buffer.empty()) {
return false; return false;
} }
llama_memory_clear(llama_get_memory(llama_ctx_v4),true);
auto res = llama_state_set_data(llama_ctx_v4, savestates[slot].current_savestate_buffer.data(), savestates[slot].current_savestate_size); auto res = llama_state_set_data(llama_ctx_v4, savestates[slot].current_savestate_buffer.data(), savestates[slot].current_savestate_size);
if(res > 0) if(res > 0)
{ {