vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/release.yml
#	README.md
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/ggml-cann.cpp
#	tools/mtmd/CMakeLists.txt
#	tools/mtmd/clip.cpp
#	tools/mtmd/clip.h
This commit is contained in:
Concedo 2025-05-24 12:10:36 +08:00
commit 55cc9acec5
35 changed files with 95818 additions and 463 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/llama-model.cpp
View file

@ -468,11 +468,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
GGML_ASSERT(hparams.n_expert_used == 0);
}
// zero-out the array hparams
std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0);
std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0);
std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
std::fill(hparams.swa_layers.begin(), hparams.swa_layers.end(), 0);
ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer, false);
ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false);
@ -579,7 +582,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.swa_type = LLAMA_SWA_TYPE_CHUNKED;
hparams.n_swa = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
hparams.n_swa_pattern = 4; // pattern: 3 chunked - 1 full
hparams.set_swa_pattern(4); // pattern: 3 chunked - 1 full
switch (hparams.n_expert) {
case 16: type = LLM_TYPE_17B_16E; break;
@ -868,7 +871,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.swa_type = LLAMA_SWA_TYPE_NONE;
hparams.n_swa = 0;
hparams.n_swa_pattern = 1;
hparams.set_swa_pattern(1);
}
} break;
case LLM_ARCH_PHIMOE:
@ -940,7 +943,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
{
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
hparams.n_swa = 4096; // default value of gemma 2
hparams.n_swa_pattern = 2;
hparams.set_swa_pattern(2);
hparams.attn_soft_cap = true;
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
@ -958,7 +961,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case LLM_ARCH_GEMMA3:
{
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
hparams.n_swa_pattern = 6;
hparams.set_swa_pattern(6);
hparams.rope_freq_base_train_swa = 10000.0f;
hparams.rope_freq_scale_train_swa = 1.0f;
@ -1043,7 +1046,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case LLM_ARCH_COHERE2:
{
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
hparams.n_swa_pattern = 4;
hparams.set_swa_pattern(4);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
@ -4417,7 +4420,7 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot);
LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa);
LLAMA_LOG_INFO("%s: n_swa_pattern = %u\n", __func__, hparams.n_swa_pattern);
LLAMA_LOG_INFO("%s: is_swa_any = %u\n", __func__, hparams.is_swa_any());
LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k);
LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v);
LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str());
@ -13289,6 +13292,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
case LLM_ARCH_JINA_BERT_V2:
case LLM_ARCH_NOMIC_BERT:
case LLM_ARCH_NOMIC_BERT_MOE:
case LLM_ARCH_WAVTOKENIZER_DEC:
{
res = nullptr;
} break;
@ -13315,7 +13319,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
GGML_ASSERT(hparams.n_swa_pattern != 1);
GGML_ASSERT(hparams.is_swa_any());
res = new llama_kv_cache_unified_iswa(
*this,
@ -13329,7 +13333,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.n_batch,
padding);
} else {
GGML_ASSERT(hparams.n_swa_pattern == 1);
GGML_ASSERT(!hparams.is_swa_any());
res = new llama_kv_cache_unified(
*this,