ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : refactor llama_context, llama_kv_cache, llm_build_context (#12181)

Browse source 
* llama : refactor llama_context, llama_kv_cache, llm_build_context

ggml-ci

* graph : don't mutate the KV cache during defrag

ggml-ci

* context : reduce virtuals + remove test function

ggml-ci

* context : move interface implementation to source file + factory

ggml-ci

* graph : move KV cache build functions to llama_context impl

ggml-ci

* graph : remove model reference from build_pooling

ggml-ci

* graph : remove llama_model reference

ggml-ci

* kv_cache : provide rope factors

ggml-ci

* graph : rework inputs to use only unique_ptr, remove attn input abstraction

ggml-ci

* context : remove llama_context_i abstraction

ggml-ci

* context : clean-up

ggml-ci

* graph : clean-up

ggml-ci

* llama : remove redundant keywords (struct, enum)

ggml-ci

* model : adapt gemma3

ggml-ci

* graph : restore same attention ops as on master

ggml-ci

* llama : remove TODO + fix indent

ggml-ci
This commit is contained in:
Georgi Gerganov 2025-03-13 12:35:44 +02:00 • committed by GitHub
parent 2048b5913d
commit e0dbec0bc6
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 13903 additions and 12190 deletions
Download patch file Download diff file Expand all files Collapse all files

24
examples/speculative/speculative.cpp
View file

@ -420,14 +420,14 @@ int main(int argc, char ** argv) {
{
LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
llama_kv_cache_seq_keep(ctx_dft, s_keep);
llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_dft, 0);
llama_kv_self_seq_keep(ctx_dft, s_keep);
llama_kv_self_seq_cp (ctx_dft, s_keep, 0, -1, -1);
llama_kv_self_seq_keep(ctx_dft, 0);
llama_kv_cache_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1);
llama_kv_cache_seq_keep(ctx_tgt, s_keep);
llama_kv_cache_seq_cp (ctx_tgt, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_tgt, 0);
llama_kv_self_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1);
llama_kv_self_seq_keep(ctx_tgt, s_keep);
llama_kv_self_seq_cp (ctx_tgt, s_keep, 0, -1, -1);
llama_kv_self_seq_keep(ctx_tgt, 0);
}
for (int s = 0; s < n_seq_dft; ++s) {
@ -444,7 +444,7 @@ int main(int argc, char ** argv) {
common_batch_clear(batch_dft);
common_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
// LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
llama_decode(ctx_dft, batch_dft);
@ -503,8 +503,8 @@ int main(int argc, char ** argv) {
if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
llama_kv_self_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
// all previous tokens from this branch are now also part of the new branch
for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@ -585,9 +585,9 @@ int main(int argc, char ** argv) {
// evaluate the target model on the drafted tokens
{
llama_kv_cache_seq_keep(ctx_tgt, 0);
llama_kv_self_seq_keep(ctx_tgt, 0);
for (int s = 1; s < n_seq_dft; ++s) {
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
}
// LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());