vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

no fast forward for empty prompt

Browse source
This commit is contained in:
Concedo 2024-07-27 16:29:35 +08:00
parent 4531ab5465
commit 729eb1e552
1 changed files with 12 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

17
gpttype_adapter.cpp
View file

@ -2335,10 +2335,14 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
} }
bool is_mamba = (file_format == FileFormat::GGUF_GENERIC && file_format_meta.model_architecture==GGUFArch::ARCH_MAMBA); bool is_mamba = (file_format == FileFormat::GGUF_GENERIC && file_format_meta.model_architecture==GGUFArch::ARCH_MAMBA);
bool blank_prompt = (addedmemory=="" && kcpp_params->prompt=="");
if (file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2 || is_mamba) if (file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2 || is_mamba)
{ {
ContextFastForward(current_context_tokens, embd_inp, n_past, last_n_tokens, nctx, smartcontext, false, true); if(!blank_prompt)
{
ContextFastForward(current_context_tokens, embd_inp, n_past, last_n_tokens, nctx, smartcontext, false, true);
}
if(is_mamba) if(is_mamba)
{ {
if(n_past==0) if(n_past==0)
@ -2355,12 +2359,15 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
else else
{ {
bool triggersc = useSmartContext; bool triggersc = useSmartContext;
if(useContextShift && (file_format == FileFormat::GGUF_GENERIC)) if(!blank_prompt) //special case for blank prompts, no fast forward or shifts
{ {
PurgeMissingTokens(llama_ctx_v4, current_context_tokens, embd_inp, inputs.max_length, nctx); if(useContextShift && (file_format == FileFormat::GGUF_GENERIC))
triggersc = false; {
PurgeMissingTokens(llama_ctx_v4, current_context_tokens, embd_inp, inputs.max_length, nctx);
triggersc = false;
}
ContextFastForward(current_context_tokens, embd_inp, n_past, last_n_tokens, nctx, smartcontext, triggersc, false);
} }
ContextFastForward(current_context_tokens, embd_inp, n_past, last_n_tokens, nctx, smartcontext, triggersc, false);
if(file_format == FileFormat::GGUF_GENERIC) if(file_format == FileFormat::GGUF_GENERIC)
{ {
llama_kv_cache_seq_rm(llama_ctx_v4, 0, n_past, -1); llama_kv_cache_seq_rm(llama_ctx_v4, 0, n_past, -1);