tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 12:05:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : fix shift

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-09 10:53:26 +03:00
parent 91a8ee6a6f
commit eee8d481d9
2 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-kv-cache-unified.cpp
View File

@ -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_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v; //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); 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); ggml_set_input(inp->k_shift);
for (const auto & layer : layers) { for (const auto & layer : layers) {

5
src/llama-kv-cells.h
View File

@ -317,8 +317,6 @@ public:
pos[i] += d; pos[i] += d;
shift[i] += d; shift[i] += d;
seq_pos_add(i);
has_shift = true; has_shift = true;
if (pos[i] < 0) { if (pos[i] < 0) {
@ -326,12 +324,15 @@ public:
seq[i].reset(); seq[i].reset();
pos[i] = -1; pos[i] = -1;
shift[i] = 0;
used.erase(i); used.erase(i);
return true; return true;
} }
seq_pos_add(i);
return false; return false;
} }