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

llama-quant : fix the verification of attention layers for encoder-decoder models (#16023)

Browse Source 
Signed-off-by: Jie Fu <jiefu@tencent.com>
This commit is contained in:
Jie Fu (傅杰)
2025-09-17 15:30:55 +08:00
committed by GitHub
parent 1cbd80f8cf
commit 745cbcf2fe
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-quant.cpp
View File

@@ -725,7 +725,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
// attention layers have a non-zero number of kv heads
int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
if (llama_model_has_encoder(&model)) {
n_attn_layer *= 3;
// now n_attn_layer is the number of attention layers in the encoder
// for each decoder block, there are 2 attention layers
n_attn_layer += 2 * model.hparams.dec_n_layer;
}
GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
}