vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

rwkv speed enhancements (batch processing), fixed a rwkv token processing bug

Browse source
This commit is contained in:
Concedo 2023-06-13 16:02:12 +08:00
parent 860fb026df
commit 9b6c35b651
2 changed files with 45 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

12
gpttype_adapter.cpp
View file

@ -409,7 +409,6 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
} }
else if (file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2) else if (file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2)
{ {
n_batch = 1;
std::string word; std::string word;
read_rwkv_vocab(); read_rwkv_vocab();
int vocabsiz = rwkv_vocab.size(); int vocabsiz = rwkv_vocab.size();
@ -425,6 +424,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
if (file_format == FileFormat::RWKV_1) if (file_format == FileFormat::RWKV_1)
{ {
n_batch = 1;
rwkv_ctx_v2 = rwkv_v2_init_from_file(modelname.c_str(), n_threads); rwkv_ctx_v2 = rwkv_v2_init_from_file(modelname.c_str(), n_threads);
//setup buffers for rwkv state //setup buffers for rwkv state
@ -453,6 +453,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
} }
else else
{ {
n_batch = 10; //use sequence mode to speedup
rwkv_ctx_v3 = rwkv_init_from_file(modelname.c_str(), n_threads); rwkv_ctx_v3 = rwkv_init_from_file(modelname.c_str(), n_threads);
//setup buffers for rwkv state //setup buffers for rwkv state
@ -946,6 +947,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
if(embd_inp.size()==0 && current_context_tokens.size()>0) if(embd_inp.size()==0 && current_context_tokens.size()>0)
{ {
embd_inp.push_back(current_context_tokens[current_context_tokens.size()-1]); embd_inp.push_back(current_context_tokens[current_context_tokens.size()-1]);
current_context_tokens.pop_back();
} }
} }
} }
@ -1014,8 +1016,16 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
rwkv_ctx_v2->state_in = rwkv_ctx_v2->state_out; rwkv_ctx_v2->state_in = rwkv_ctx_v2->state_out;
} }
else else
{
if(embd.size()>1)
{
evalres = rwkv_eval_sequence(rwkv_ctx_v3, (uint32_t*)embd.data(), embd.size(), rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
}
else
{ {
evalres = rwkv_eval(rwkv_ctx_v3, embd[0], rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out); evalres = rwkv_eval(rwkv_ctx_v3, embd[0], rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
}
memcpy(logits.data(), rwkv_ctx_v3->logits_out, sizeof(float) * rwkv_vocab.size()); memcpy(logits.data(), rwkv_ctx_v3->logits_out, sizeof(float) * rwkv_vocab.size());
rwkv_ctx_v3->state_in = rwkv_ctx_v3->state_out; rwkv_ctx_v3->state_in = rwkv_ctx_v3->state_out;
} }