vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-08 01:41:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge commit '2b6dfe824d' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/release.yml
#	examples/save-load-state/save-load-state.cpp
#	src/llama-context.cpp
#	tools/cli/cli.cpp
This commit is contained in:
Concedo 2026-02-26 15:07:23 +08:00
commit 7e53bfd28d
17 changed files with 258 additions and 194 deletions
Download patch file Download diff file Expand all files Collapse all files

62
common/common.cpp
View file

@ -1767,3 +1767,65 @@ float lr_opt::get_lr(float epoch) const {
LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
return r;
}
bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos) {
llama_batch batch = llama_batch_get_one(&last_token, 1);
batch.pos = &pos;
if (llama_decode(ctx, batch)) {
LOG_ERR("%s: failed to replay last token\n", __func__);
return false;
}
return true;
}
bool common_prompt_batch_decode(
struct llama_context * ctx,
const std::vector<llama_token> & tokens,
int & n_past,
int n_batch,
std::string_view state_path,
bool save_state) {
const int n_eval = tokens.size();
if (n_eval == 0) {
return true;
}
if (save_state && n_eval > 1) {
const int n_tokens_before_last = n_eval - 1;
GGML_ASSERT(n_eval <= n_batch);
// Decode all but the last token so we can save the memory state before decoding the last token.
// This is done so we can restore the session state later and replay the last token.
// Memory implementations in recurrent/hybrid models don't support removing tokens from their
// memory, so we can't just remove the last token from the memory and replay the last token which
// is the reason for this logic.
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_tokens_before_last))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
n_past += n_tokens_before_last;
llama_state_save_file(ctx, state_path.data(), tokens.data(), n_tokens_before_last);
LOG_INF("saved session before last token to %s, n_tokens = %d\n", state_path.data(), n_tokens_before_last);
llama_token last_token = tokens.back();
llama_batch batch = llama_batch_get_one(&last_token, 1);
int32_t pos = n_past;
batch.pos = &pos;
if (llama_decode(ctx, batch)) {
LOG_ERR("%s : failed to eval last token\n", __func__);
return false;
}
n_past++;
} else {
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
n_past += n_eval;
}
return true;
}