tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : fix shift and defrag logic (#14081)

Browse Source 
* kv-cache : fix shift

ggml-ci

* cont : reset shift[i]

ggml-ci

* cont : fix defrag erasing cells that didn't move

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-09 23:04:35 +03:00
committed by GitHub
parent 7f4fbe5183
commit 40cbf571c9
2 changed files with 12 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@ -462,7 +462,7 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
for (uint32_t i = 0; i < n_kv; ++i) {
assert(dinfo.ids[i] <= n_kv);
if (dinfo.ids[i] == n_kv) {
if (dinfo.ids[i] == n_kv || dinfo.ids[i] == i) {
continue;
}
@ -944,11 +944,9 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
const auto & n_embd_head_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v;
//GGML_ASSERT(kv_self->size == n_ctx);
auto inp = std::make_unique<llm_graph_input_k_shift>(this);
inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cparams.n_ctx);
inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cells.size());
ggml_set_input(inp->k_shift);
for (const auto & layer : layers) {