upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-05 09:30:49 -05:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : improve find_slot impl

Browse Source
This commit is contained in:
Georgi Gerganov
2025-07-02 13:37:12 +03:00
parent 2ac5be3a58
commit a70293bc25
1 changed files with 7 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/llama-kv-cache-unified.cpp
View File

@@ -624,14 +624,15 @@ llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_
}
}
uint32_t n_found = 0;
uint32_t n_tested = 0;
const uint32_t n_test = cont ? n_tokens : 1;
slot_info res;
res.idxs.resize(n_tokens);
auto & idxs = res.idxs;
idxs.reserve(n_tokens);
while (true) {
if (head_cur + n_test > cells.size()) {
@@ -677,20 +678,18 @@ llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_
n_tested++;
if (can_use) {
res.idxs[n_found] = idx;
n_found++;
idxs.push_back(idx);
} else {
break;
}
}
if (n_found == n_tokens) {
if (idxs.size() == n_tokens) {
break;
}
if (cont) {
n_found = 0;
idxs.clear();
}
if (n_tested >= cells.size()) {
@@ -700,7 +699,7 @@ llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_
}
// we didn't find a suitable slot - return empty result
if (n_found < n_tokens) {
if (idxs.size() < n_tokens) {
res.clear();
}