vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-17 21:20:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge commit 'ddf9f94389' into concedo_experimental

Browse source 
# Conflicts:
#	examples/model-conversion/scripts/causal/run-converted-model.sh
#	examples/model-conversion/scripts/causal/run-org-model.py
#	src/CMakeLists.txt
#	src/llama-quant.cpp
#	tools/server/README.md
This commit is contained in:
Concedo 2025-11-28 23:27:50 +08:00
commit 0ccb298087
23 changed files with 2813 additions and 86 deletions
Download patch file Download diff file Expand all files Collapse all files

18
src/llama-quant.cpp
View file

@ -684,7 +684,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
}
LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
continue;
} else if (remapped_name != it.first) {
}
if (remapped_name != it.first) {
ggml_set_name(it.second.tensor, remapped_name.c_str());
LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
}
@ -729,13 +731,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
{
const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
// 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);
int32_t n_layer_attn = 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)) {
// now n_attn_layer is the number of attention layers in the encoder
// now n_layer_attn 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;
n_layer_attn += 2 * model.hparams.dec_n_layer;
}
GGML_ASSERT_CONTINUE((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
// note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
GGML_ASSERT_CONTINUE((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
}
size_t total_size_org = 0;