Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

another fix for drafting

Browse Source
This commit is contained in:
Concedo
2026-06-23 22:34:34 +08:00
parent 6df4ca13f1
commit 19064083bd
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpttype_adapter.cpp
+5
View File
@@ -2929,6 +2929,11 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
llama_ctx_params.offload_kqv = !inputs.low_vram;
llama_ctx_params.kv_unified = true;
if((inputs.use_mtp || draftmodel_filename != "") && inputs.draft_amount > 0)
{
// Match llama-server's target rollback slots for speculative verification.
llama_ctx_params.n_rs_seq = inputs.draft_amount;
}
model_params.use_mmap = inputs.use_mmap;
model_params.use_mlock = inputs.use_mlock;
model_params.use_direct_io = false; //no direct io for now until stable