vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-11 13:11:49 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build-linux-cross.yml
#	.github/workflows/build.yml
#	CODEOWNERS
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cuda/fattn.cu
#	ggml/src/ggml-webgpu/CMakeLists.txt
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tools/llama-bench/llama-bench.cpp
#	tools/rpc/README.md
#	tools/server/README.md
This commit is contained in:
Concedo 2025-10-09 01:33:27 +08:00
commit b6f6338bba
32 changed files with 1556 additions and 636 deletions
Download patch file Download diff file Expand all files Collapse all files

28
src/llama-arch.cpp
View file

@ -93,6 +93,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
{ LLM_ARCH_LFM2, "lfm2" },
{ LLM_ARCH_LFM2MOE, "lfm2moe" },
{ LLM_ARCH_DREAM, "dream" },
{ LLM_ARCH_SMALLTHINKER, "smallthinker" },
{ LLM_ARCH_LLADA, "llada" },
@ -2104,6 +2105,32 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT, "output" },
}
},
{
LLM_ARCH_LFM2MOE,
{
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" },
{ LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" },
{ LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
}
},
{
LLM_ARCH_SMALLTHINKER,
{
@ -2493,6 +2520,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_PLAMO2:
case LLM_ARCH_GRANITE_HYBRID:
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_NEMOTRON_H:
return true;
default:

1
src/llama-arch.h
View file

@ -97,6 +97,7 @@ enum llm_arch {
LLM_ARCH_SMOLLM3,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_LFM2,
LLM_ARCH_LFM2MOE,
LLM_ARCH_DREAM,
LLM_ARCH_SMALLTHINKER,
LLM_ARCH_LLADA,

2
src/llama-chat.cpp
View file

@ -590,7 +590,7 @@ int32_t llm_chat_apply_template(
ss << message->content << "<|end_of_text|>\n";
}
if (add_ass) {
ss << "<|start_of_role|>assistant<|end_of_role|>\n";
ss << "<|start_of_role|>assistant<|end_of_role|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
// GigaChat template

4
src/llama-memory-hybrid.cpp
View file

@ -73,7 +73,9 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch, false);
// TODO: non-sequential equal split can be done if using unified KV cache
// for simplicity, we always use sequential equal split for now
ubatch = balloc.split_equal(n_ubatch, true);
}
if (ubatch.n_tokens == 0) {

9
src/llama-memory-recurrent.cpp
View file

@ -382,7 +382,9 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch, false);
// TODO: non-sequential equal split can be done if using unified KV cache
// for simplicity, we always use sequential equal split for now
ubatch = balloc.split_equal(n_ubatch, true);
}
if (ubatch.n_tokens == 0) {
@ -859,9 +861,12 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
if (dest_seq_id != -1) {
// single sequence
seq_rm(dest_seq_id, -1, -1);
if (cell_count == 0) {
return true;
}
llama_batch_allocr balloc(hparams.n_pos_per_embd());
llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);

98
src/llama-model.cpp
View file

@ -119,6 +119,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_17B_16E: return "17Bx16E (Scout)";
case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)";
case LLM_TYPE_A13B: return "A13B";
case LLM_TYPE_8B_A1B: return "8B.A1B";
case LLM_TYPE_21B_A3B: return "21B.A3B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_106B_A12B: return "106B.A12B";
@ -315,7 +316,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
}
// CPU: ACCEL -> GPU host -> CPU extra -> CPU
static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts) {
static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts, bool no_host) {
buft_list_t buft_list;
// add ACCEL buffer types
@ -336,11 +337,13 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
// generally, this will be done using the first device in the list
// a better approach would be to handle this on a weight-by-weight basis using the offload_op
// function of the device to determine if it would benefit from being stored in a host buffer
for (auto * dev : devices) {
ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
if (buft) {
buft_list.emplace_back(dev, buft);
break;
if (!no_host) {
for (auto * dev : devices) {
ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
if (buft) {
buft_list.emplace_back(dev, buft);
break;
}
}
}
@ -1998,14 +2001,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
}
hparams.n_layer_dense_lead = hparams.n_layer;
switch (hparams.n_ff()) {
case 4608: type = LLM_TYPE_350M; break;
case 6912: type = LLM_TYPE_700M; break;
case 8192: type = LLM_TYPE_1_2B; break;
case 10752: type = LLM_TYPE_2_6B; break;
default: type = LLM_TYPE_UNKNOWN;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_LFM2MOE:
{
ml.get_key(LLM_KV_SHORTCONV_L_CACHE, hparams.n_shortconv_l_cache);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
}
type = LLM_TYPE_8B_A1B;
} break;
case LLM_ARCH_SMALLTHINKER:
{
const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
@ -2088,7 +2106,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
// build a list of buffer types for the CPU and GPU devices
pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts);
pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
for (auto * dev : devices) {
buft_list_t buft_list = make_gpu_buft_list(dev, split_mode, tensor_split);
// add CPU buffer types as a fallback
@ -5870,6 +5888,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
@ -5881,11 +5900,23 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
// ffn is same for transformer and conv layers
const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);
// ffn/moe is same for transformer and conv layers
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
if (is_moe_layer) {
GGML_ASSERT(n_expert && n_expert_used);
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
} else { // dense
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
// for operator_norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
@ -6367,7 +6398,7 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
}
if (arch == LLM_ARCH_SMALLTHINKER) {
if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
}
@ -18662,6 +18693,8 @@ struct llm_build_lfm2 : public llm_graph_context {
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
auto * prev_cur = cur;
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "model.layers.{}.operator_norm", il);
@ -18676,7 +18709,16 @@ struct llm_build_lfm2 : public llm_graph_context {
}
cur = ggml_add(ctx0, prev_cur, cur);
cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
ggml_tensor * ffn_out = is_moe_layer ?
build_moe_feed_forward(ffn_norm_out, il) :
build_dense_feed_forward(ffn_norm_out, il);
cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_out);
}
cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
@ -18691,23 +18733,32 @@ struct llm_build_lfm2 : public llm_graph_context {
ggml_build_forward_expand(gf, cur);
}
ggml_tensor * build_feed_forward(ggml_tensor * cur,
int il) const {
cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "model.layers.{}.ffn_norm", il);
ggml_tensor * build_moe_feed_forward(ggml_tensor * cur,
int il) const {
return build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
il);
}
ggml_tensor * build_dense_feed_forward(ggml_tensor * cur,
int il) const {
GGML_ASSERT(!model.layers[il].ffn_up_b);
GGML_ASSERT(!model.layers[il].ffn_gate_b);
GGML_ASSERT(!model.layers[il].ffn_down_b);
cur = build_ffn(cur,
return build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "model.layers.{}.feed_forward.w2", il);
return cur;
}
ggml_tensor * build_attn_block(ggml_tensor * cur,
@ -19877,6 +19928,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
llm = std::make_unique<llm_build_falcon_h1>(*this, params);
} break;
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
{
llm = std::make_unique<llm_build_lfm2>(*this, params);
} break;
@ -19927,6 +19979,7 @@ llama_model_params llama_model_default_params() {
/*.use_mlock =*/ false,
/*.check_tensors =*/ false,
/*.use_extra_bufts =*/ true,
/*.no_host =*/ false,
};
return result;
@ -20098,6 +20151,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_OPENAI_MOE:
case LLM_ARCH_HUNYUAN_DENSE:
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_SMALLTHINKER:
case LLM_ARCH_GLM4_MOE:
case LLM_ARCH_SEED_OSS:

1
src/llama-model.h
View file

@ -107,6 +107,7 @@ enum llm_type {
LLM_TYPE_17B_16E, // llama4 Scout
LLM_TYPE_17B_128E, // llama4 Maverick
LLM_TYPE_A13B,
LLM_TYPE_8B_A1B, // lfm2moe
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_106B_A12B, // GLM-4.5-Air