vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-07-22 23:33:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fixed mtp for qwen

Browse source
This commit is contained in:
Concedo 2026-06-11 22:04:31 +08:00
parent 40084bb1f8
commit 79331fd382
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
gpttype_adapter.cpp
View file

@ -6657,7 +6657,7 @@ size_t gpttype_save_state_kv(int slot)
savestates[slot].current_savestate_size = newsize;
savestates[slot].savestate_context_tokens = current_context_tokens;
savestates[slot].media_signature = media_composite_image_signature;
float * lgptr = llama_get_logits(llama_ctx_v4);
float * lgptr = (draft_is_mtp ? llama_get_logits_ith(llama_ctx_v4, -1) : llama_get_logits(llama_ctx_v4));
savestates[slot].latest_logits.assign(lgptr,lgptr+n_vocab);
int maxedpos = llama_memory_seq_pos_max(llama_get_memory(llama_ctx_v4),0);
//kcpp: so maxedpos appears to always be equal to ctx tokens - 2, if savestate_ctx_tokens > maxedpos + 2 then trim excess