CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

context : fix logits size overflow for huge batches

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-08-04 22:26:34 -04:00
parent f16a843a38
commit 145401c9e3
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-context.cpp
View File

@@ -786,7 +786,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;
const int64_t n_embd = hparams.n_embd;
const int32_t n_vocab = model.vocab.n_tokens();
const int64_t n_vocab = model.vocab.n_tokens();
// note: during encode, we always pass the full sequence starting from pos = 0
if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
@@ -959,7 +959,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
const auto & vocab = model.vocab;
const auto & hparams = model.hparams;
const int32_t n_vocab = vocab.n_tokens();
const int64_t n_vocab = vocab.n_tokens();
const int64_t n_embd = hparams.n_embd;
// when computing embeddings, all tokens are output