vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-09 19:46:11 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

kv-cache : support V-less cache (#19067)

Browse source 
* kv-cache : support V-less cache

* cuda : better check for V_is_K_view

* cuda : improve V_is_K_view check

* graph : add comments

* hparams : refactor
This commit is contained in:
Georgi Gerganov 2026-01-25 15:48:56 +02:00 • committed by GitHub
parent 70d860824a
commit d9c6ce46f7
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 246 additions and 53 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/llama-context.cpp
View file

@ -793,7 +793,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
}
const uint32_t n_embd_out = model.hparams.get_n_embd_out();
const uint32_t n_embd_out = model.hparams.n_embd_out();
return embd + j*n_embd_out;
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
@ -1279,7 +1279,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
{
// extract token embeddings
GGML_ASSERT(embd != nullptr);
const uint32_t n_embd_out = hparams.get_n_embd_out();
const uint32_t n_embd_out = hparams.n_embd_out();
GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
@ -1688,7 +1688,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
{
// extract token embeddings
GGML_ASSERT(embd != nullptr);
const uint32_t n_embd_out = hparams.get_n_embd_out();
const uint32_t n_embd_out = hparams.n_embd_out();
float * embd_out = embd + n_outputs_prev*n_embd_out;
if (n_outputs) {
@ -1821,7 +1821,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
const auto n_batch = cparams.n_batch;
const auto n_vocab = vocab.n_tokens();
const auto n_embd_out = hparams.get_n_embd_out();
const auto n_embd_out = hparams.n_embd_out();
bool has_logits = true;
bool has_embd = cparams.embeddings;