thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

kv-cache : fix M-RoPE checkpoints (#20132)

Browse source
This commit is contained in:
Georgi Gerganov 2026-03-06 08:46:51 +02:00 • committed by GitHub
parent f7db3f3789
commit 17a4258946
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 15 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama-batch.cpp
View file

@ -394,11 +394,13 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
clear();
split_reset();
const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd;
auto udata = std::make_shared<llama_ubatch::data_t>();
udata->token .resize(n_tokens);
udata->embd .clear();
udata->pos .resize(n_tokens);
udata->pos .resize(n_pos_all);
udata->n_seq_id .resize(n_tokens);
udata->seq_id .resize(n_tokens);
udata->seq_id_unq.resize(0);