wylab/llama.cpp
1
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-06-09 07:16:44 +02:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache: follow the source cache size when sharing cells (#24267)

Browse Source 
A fitted target context can end up smaller than the draft default, the
oversized assistant views then overflow the shared K/V tensors and trip
the ggml_view_4d size assert during graph reserve.
This commit is contained in:
Pascal
2026-06-07 17:33:00 +02:00
committed by GitHub
parent 04eb4c446d
commit f0156d1401
1 changed files with 11 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-kv-cache.cpp
+11
View File
@@ -97,6 +97,17 @@ llama_kv_cache::llama_kv_cache(
model(model), hparams(hparams), v_trans(v_trans),
n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
// shared cells view the source cache's K/V tensors, so the cell count
// follows the source allocation: a fitted target can be smaller than the
// draft default and oversized views would overflow the source tensors
if (mem_other) {
const uint32_t size_other = static_cast<llama_kv_cache *>(mem_other)->get_size();
if (kv_size != size_other) {
LLAMA_LOG_WARN("%s: kv_size = %u overridden to %u to match the shared source cache\n", __func__, kv_size, size_other);
kv_size = size_other;
}
}
GGML_ASSERT(kv_size % n_pad == 0);
const uint32_t n_layer = hparams.n_layer_all;