Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	ggml/src/ggml-cuda/ggml-cuda.cu
#	tests/test-backend-ops.cpp
#	tools/mtmd/CMakeLists.txt
This commit is contained in:
Concedo
2026-01-02 11:05:20 +08:00
parent 0a23388e7d e86f3c2221
commit 7e1ae49e7d
33 changed files with 921 additions and 248 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-chat.cpp
+11
View File
@@ -74,6 +74,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "seed_oss", LLM_CHAT_TEMPLATE_SEED_OSS },
{ "grok-2", LLM_CHAT_TEMPLATE_GROK_2 },
{ "pangu-embedded", LLM_CHAT_TEMPLATE_PANGU_EMBED },
{ "solar-open", LLM_CHAT_TEMPLATE_SOLAR_OPEN },
};
llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -216,6 +217,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_GROK_2;
} else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
return LLM_CHAT_TEMPLATE_PANGU_EMBED;
} else if (tmpl_contains("<|begin|>") && tmpl_contains("<|end|>") && tmpl_contains("<|content|>")) {
return LLM_CHAT_TEMPLATE_SOLAR_OPEN;
}
return LLM_CHAT_TEMPLATE_UNKNOWN;
}
@@ -845,6 +848,14 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "[unused9]助手：";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_SOLAR_OPEN) {
for (auto message : chat) {
std::string role(message->role);
ss << "<|begin|>" << role << "<|content|>" << message->content << "<|end|>";
}
if (add_ass) {
ss << "<|begin|>assistant";
}
} else {
// template not supported
return -1;
src/llama-chat.h
+1
View File
@@ -54,6 +54,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_SEED_OSS,
LLM_CHAT_TEMPLATE_GROK_2,
LLM_CHAT_TEMPLATE_PANGU_EMBED,
LLM_CHAT_TEMPLATE_SOLAR_OPEN,
LLM_CHAT_TEMPLATE_UNKNOWN,
};
src/llama-model.cpp
+25 -8
View File
@@ -234,6 +234,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_31B_A3_5B: return "31B.A3.5B";
case LLM_TYPE_80B_A3B: return "80B.A3B";
case LLM_TYPE_100B_A6B: return "100B.A6B";
case LLM_TYPE_102B_A12B: return "102B.A12B";
case LLM_TYPE_106B_A12B: return "106B.A12B";
case LLM_TYPE_230B_A10B: return "230B.A10B";
case LLM_TYPE_235B_A22B: return "235B.A22B";
@@ -1790,7 +1791,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
@@ -1886,6 +1887,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
switch (hparams.n_layer) {
case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
case 48: type = LLM_TYPE_102B_A12B; break; // Solar Open
case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
default: type = LLM_TYPE_UNKNOWN;
}
@@ -3481,7 +3483,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, layer.ffn_gate ? n_ff : n_ff * 2}, 0);
const auto tn_ffn_up_weight = tn(LLM_TENSOR_FFN_UP, "weight", i);
ggml_tensor * t_ffn_up = ml.get_tensor_meta(tn_ffn_up_weight.str().c_str());
const int64_t n_ffn_up = t_ffn_up ? t_ffn_up->ne[1] : n_ff;
GGML_ASSERT(n_ffn_up == n_ff || n_ffn_up == n_ff * 2);
layer.ffn_up = create_tensor(tn_ffn_up_weight, {n_embd, n_ffn_up}, 0);
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ffn_up}, TENSOR_NOT_REQUIRED);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
@@ -4937,7 +4946,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
// try to load output.weight, if not found, use token_embd (tied embeddings)
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
if (!output) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
@@ -5000,7 +5013,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
// try to load output.weight, if not found, use token_embd (tied embeddings)
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
if (!output) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
@@ -5367,9 +5384,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, flags);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, flags);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, flags);
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, flags);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, flags);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, flags);
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, TENSOR_NOT_REQUIRED | flags);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, TENSOR_NOT_REQUIRED | flags);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, TENSOR_NOT_REQUIRED | flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
@@ -7605,7 +7622,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
} break;
case LLM_ARCH_MODERN_BERT:
{
llm = std::make_unique<llm_build_modern_bert<true>>(*this, params);
llm = std::make_unique<llm_build_modern_bert>(*this, params);
} break;
case LLM_ARCH_NEO_BERT:
{
src/llama-model.h
+1
View File
@@ -119,6 +119,7 @@ enum llm_type {
LLM_TYPE_31B_A3_5B,
LLM_TYPE_80B_A3B, // Qwen3 Next
LLM_TYPE_100B_A6B,
LLM_TYPE_102B_A12B, // Solar-Open
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_230B_A10B, // Minimax M2
LLM_TYPE_235B_A22B,
src/llama-vocab.cpp
+26 -5
View File
@@ -539,6 +539,12 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_YOUTU:
regex_exprs = {
"[가-힣ㄱ-ㆎ]+|[！…“”‘’—：；，、-〿︰-﹏]+|[ㄅ-ㄯ]+|[一-龥぀-ゟ゠-ヿ]+",
"[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
regex_exprs = {
"[\r\n]",
@@ -580,6 +586,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
case LLAMA_VOCAB_PRE_TYPE_QWEN2:
case LLAMA_VOCAB_PRE_TYPE_HUNYUAN:
case LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
@@ -2096,6 +2103,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "deepseek-v3") {
pre_type = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM;
clean_spaces = false;
} else if (
tokenizer_pre == "youtu") {
pre_type = LLAMA_VOCAB_PRE_TYPE_YOUTU;
clean_spaces = false;
ignore_merges = true;
} else if (
tokenizer_pre == "falcon") {
pre_type = LLAMA_VOCAB_PRE_TYPE_FALCON;
@@ -2251,6 +2263,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "minimax-m2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
clean_spaces = false;
} else if (
tokenizer_pre == "solar-open") {
pre_type = LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN;
clean_spaces = false;
} else {
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
}
@@ -2597,6 +2613,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|end|>"
|| t.first == "<|return|>" // o200k_harmony
|| t.first == "<|call|>" // o200k_harmony
|| t.first == "<|flush|>" // solar-open
|| t.first == "<|calls|>" // solar-open
|| t.first == "<end_of_turn>"
|| t.first == "<|endoftext|>"
|| t.first == "<|eom_id|>"
@@ -2643,13 +2661,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
}
// TODO: workaround for o200k_harmony tokenizer: the "<|end|>" token should not be EOG
// we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens,
// TODO: workaround for o200k_harmony and solar-open tokenizer: the "<|end|>" token should not be EOG
// we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens ("<|calls|>" and "<|flush|>" for solar-open),
// we remove the "<|end|>" token from the EOG list
{
bool has_return = false;
bool has_call = false;
bool has_end = false;
bool has_flush = false;
llama_token end_id = LLAMA_TOKEN_NULL;
@@ -2659,18 +2678,20 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (id_to_token[tid].text == "<|return|>") {
has_return = true;
} else if (id_to_token[tid].text == "<|call|>") {
} else if (id_to_token[tid].text == "<|call|>" || id_to_token[tid].text == "<|calls|>") {
has_call = true;
} else if (id_to_token[tid].text == "<|flush|>") {
has_flush = true;
} else if (id_to_token[tid].text == "<|end|>") {
has_end = true;
end_id = tid;
}
}
if (has_return && has_call && has_end) {
if ((has_return && has_call && has_end) || (has_call && has_flush && has_end)) {
special_eog_ids.erase(end_id);
id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>', or '<|calls|>' and '<|flush|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
}
}
}
src/llama-vocab.h
+2
View File
@@ -52,6 +52,8 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2 = 41,
LLAMA_VOCAB_PRE_TYPE_AFMOE = 42,
LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43,
LLAMA_VOCAB_PRE_TYPE_YOUTU = 44,
};
struct LLM_KV;
src/models/bert.cpp
+4 -2
View File
@@ -142,11 +142,13 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
LLM_FFN_GELU, LLM_FFN_SEQ, il);
cb(cur, "ffn_out", il);
} else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();
auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
type_op, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
cur = build_ffn(cur,
src/models/deepseek2.cpp
+1 -1
View File
@@ -215,7 +215,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/models.h
-1
View File
@@ -332,7 +332,6 @@ struct llm_build_mistral3 : public llm_graph_context {
llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
};
template <bool iswa>
struct llm_build_modern_bert : public llm_graph_context {
llm_build_modern_bert(const llama_model & model, const llm_graph_params & params);
};
src/models/modern-bert.cpp
+2 -13
View File
@@ -1,7 +1,6 @@
#include "models.h"
template <bool iswa>
llm_build_modern_bert<iswa>::llm_build_modern_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
@@ -24,13 +23,7 @@ llm_build_modern_bert<iswa>::llm_build_modern_bert(const llama_model & model, co
auto * inp_attn = build_attn_inp_no_cache();
for (int il = 0; il < n_layer; ++il) {
float freq_base_l = 0.0f;
if constexpr (iswa) {
freq_base_l = model.get_rope_freq_base(cparams, il);
} else {
freq_base_l = freq_base;
}
float freq_base_l = model.get_rope_freq_base(cparams, il);
cur = inpL;
@@ -120,7 +113,3 @@ llm_build_modern_bert<iswa>::llm_build_modern_bert(const llama_model & model, co
res->t_embd = cur;
ggml_build_forward_expand(gf, cur);
}
// Explicit template instantiations
template struct llm_build_modern_bert<false>;
template struct llm_build_modern_bert<true>;
src/unicode.cpp
+23 -14
View File
@@ -972,6 +972,11 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
{ "\\p{P}", unicode_cpt_flags::PUNCTUATION },
{ "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
{ "\\p{S}", unicode_cpt_flags::SYMBOL },
{ "\\p{Lu}", unicode_cpt_flags::LETTER }, // Uppercase letter
{ "\\p{Ll}", unicode_cpt_flags::LETTER }, // Lowercase letter
{ "\\p{Lt}", unicode_cpt_flags::LETTER }, // Titlecase letter
{ "\\p{Lm}", unicode_cpt_flags::LETTER }, // Modifier letter
{ "\\p{Lo}", unicode_cpt_flags::LETTER }, // Other letter
};
static const std::map<int, int> k_ucat_cpt = {
@@ -1082,22 +1087,26 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
continue;
}
if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
// Match \p{...} Unicode properties of varying lengths
if (regex_expr[i + 0] == '\\' && i + 3 < regex_expr.size() &&
regex_expr[i + 1] == 'p' &&
regex_expr[i + 2] == '{' &&
regex_expr[i + 4] == '}') {
const std::string pat = regex_expr.substr(i, 5);
if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
if (!inside) {
regex_expr_collapsed += '[';
regex_expr[i + 2] == '{') {
// Find the closing brace
size_t closing_brace = regex_expr.find('}', i + 3);
if (closing_brace != std::string::npos && closing_brace <= i + 10) { // reasonable limit
const std::string pat = regex_expr.substr(i, closing_brace - i + 1);
if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
if (!inside) {
regex_expr_collapsed += '[';
}
regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
if (!inside) {
regex_expr_collapsed += ']';
}
i = closing_brace;
continue;
}
regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
if (!inside) {
regex_expr_collapsed += ']';
}
i += 4;
continue;
}
}