CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

cont : restore padding for n_kv tensor shape

Browse Source
This commit is contained in:
Georgi Gerganov
2025-10-28 11:09:59 +02:00
parent e28cec364e
commit e776168267
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llama-kv-cache.cpp
View File

@@ -957,10 +957,14 @@ bool llama_kv_cache::get_has_shift() const {
uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const { uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
uint32_t result = 0; uint32_t result = 0;
// pad the n_kv value so that the graph remains constant across batches and can be reused
// note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
const uint32_t n_pad_cur = std::max(n_pad, 256u);
for (uint32_t s = 0; s < sinfo.n_stream(); ++s) { for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
const auto & cells = v_cells[sinfo.strm[s]]; const auto & cells = v_cells[sinfo.strm[s]];
result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result); result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
} }
return result; return result;