vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-18 23:49:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	README.md
#	docs/backend/zDNN.md
#	docs/build.md
#	docs/ops.md
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-rpc/ggml-rpc.cpp
#	ggml/src/ggml-sycl/convert.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	src/llama-quant.cpp
#	tests/test-backend-ops.cpp
#	tools/llama-bench/llama-bench.cpp
#	tools/server/README.md
This commit is contained in:
Concedo 2025-12-07 16:48:38 +08:00
commit 17c0c8d55d
40 changed files with 1161 additions and 326 deletions
Download patch file Download diff file Expand all files Collapse all files

27
src/llama-quant.cpp
View file

@ -669,7 +669,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
std::map<int, std::string> mapped;
int blk_id = 0;
int pruned_attention_w = 0;
// make a list of weights
std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
@ -677,11 +676,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
for (const auto & it : ml.weights_map) {
const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
if (remapped_name.empty()) {
if (it.first.find("attn_v.weight") != std::string::npos ||
it.first.find("attn_qkv.weight") != std::string::npos ||
it.first.find("attn_kv_b.weight") != std::string::npos) {
pruned_attention_w++;
}
LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
continue;
}
@ -706,7 +700,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
});
}
bool is_clip_model = false;
for (const auto * it : tensors) {
const struct ggml_tensor * tensor = it->tensor;
@ -720,30 +713,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
} else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
qs.has_output = true;
}
is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
}
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
// sanity checks for models that have attention layers
if (qs.n_attention_wv != 0 && !is_clip_model)
{
int32_t n_layer_all = model.hparams.n_layer;
if (llama_model_has_encoder(&model)) {
// now n_layer_all is the number of attention layers in the encoder
// for each decoder block, there are 2 attention layers
n_layer_all += 2 * model.hparams.dec_n_layer;
}
// 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_all = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_all, n_layer_recr, pruned_attention_w);
GGML_ASSERT_CONTINUE((qs.n_attention_wv == n_layer_all - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
}
size_t total_size_org = 0;
size_t total_size_new = 0;