Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

calculate some fields before autofit for more accurate estimate

Browse Source
This commit is contained in:
Concedo
2026-04-06 20:44:37 +08:00
parent f6e712d919
commit 82cc19e055
1 changed files with 7 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpttype_adapter.cpp
+7 -5
View File
@@ -2548,6 +2548,13 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
model_params.tensor_buft_overrides = tenos.data();
}
//set some ctx params early so autofit can use them.
llama_ctx_params.flash_attn_type = (kcpp_data->flash_attn?LLAMA_FLASH_ATTN_TYPE_ENABLED:LLAMA_FLASH_ATTN_TYPE_DISABLED);
llama_ctx_params.swa_full = kcpp_data->swa_full;
llama_ctx_params.type_k = (inputs.quant_k==2?GGML_TYPE_Q4_0:(inputs.quant_k==1?GGML_TYPE_Q8_0:(inputs.quant_k==3?GGML_TYPE_BF16:GGML_TYPE_F16)));
llama_ctx_params.type_v = (inputs.quant_v==2?GGML_TYPE_Q4_0:(inputs.quant_v==1?GGML_TYPE_Q8_0:(inputs.quant_v==3?GGML_TYPE_BF16:GGML_TYPE_F16)));
//apply overrides from autofit
float tensor_split_temp[128] = {0}; //temp buffer for autofit
std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(),1024*1024*1024);
@@ -2666,11 +2673,6 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
llamamodel->vocab.set_eos_bos(0,0);
}
llama_ctx_params.flash_attn_type = (kcpp_data->flash_attn?LLAMA_FLASH_ATTN_TYPE_ENABLED:LLAMA_FLASH_ATTN_TYPE_DISABLED);
llama_ctx_params.swa_full = kcpp_data->swa_full;
llama_ctx_params.type_k = (inputs.quant_k==2?GGML_TYPE_Q4_0:(inputs.quant_k==1?GGML_TYPE_Q8_0:(inputs.quant_k==3?GGML_TYPE_BF16:GGML_TYPE_F16)));
llama_ctx_params.type_v = (inputs.quant_v==2?GGML_TYPE_Q4_0:(inputs.quant_v==1?GGML_TYPE_Q8_0:(inputs.quant_v==3?GGML_TYPE_BF16:GGML_TYPE_F16)));
llama_ctx_v4 = llama_init_from_model(llamamodel, llama_ctx_params);
if(load_guidance)
{