vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-09 02:50:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build.yml
#	ggml/CMakeLists.txt
#	ggml/cmake/ggml-config.cmake.in
#	ggml/src/CMakeLists.txt
#	models/templates/README.md
#	tools/imatrix/imatrix.cpp
This commit is contained in:
Concedo 2025-08-05 19:24:29 +08:00
commit 7590a0ea39
22 changed files with 1122 additions and 421 deletions
Download patch file Download diff file Expand all files Collapse all files

44
src/llama-arch.cpp
View file

@ -62,6 +62,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
{ LLM_ARCH_BITNET, "bitnet" },
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_T5ENCODER, "t5encoder" },
@ -127,6 +128,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
{ LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
{ LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
{ LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
{ LLM_KV_POOLING_TYPE, "%s.pooling_type" },
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
{ LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
@ -1391,6 +1393,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
},
},
{
LLM_ARCH_GLM4_MOE,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_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_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ 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_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
// NextN/MTP tensors - preserved but unused (in final layer, dynamic layer number)
{ LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" },
{ LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
{ LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" },
{ LLM_TENSOR_NEXTN_HNORM, "blk.%d.nextn.hnorm" },
{ LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
{ LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
},
},
{
LLM_ARCH_BITNET,
{
@ -2181,6 +2217,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
{LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
// NextN/MTP tensors are currently ignored (reserved for future MTP support)
// These tensors only exist in the last layer(s) and are treated as output tensors
{LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
};
LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

8
src/llama-arch.h
View file

@ -66,6 +66,7 @@ enum llm_arch {
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
LLM_ARCH_BITNET,
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
@ -131,6 +132,7 @@ enum llm_kv {
LLM_KV_EXPERT_WEIGHTS_NORM,
LLM_KV_EXPERT_GATING_FUNC,
LLM_KV_MOE_EVERY_N_LAYERS,
LLM_KV_NEXTN_PREDICT_LAYERS,
LLM_KV_POOLING_TYPE,
LLM_KV_LOGIT_SCALE,
LLM_KV_DECODER_START_TOKEN_ID,
@ -409,6 +411,12 @@ enum llm_tensor {
LLM_TENSOR_SHORTCONV_CONV,
LLM_TENSOR_SHORTCONV_INPROJ,
LLM_TENSOR_SHORTCONV_OUTPROJ,
LLM_TENSOR_NEXTN_EH_PROJ,
LLM_TENSOR_NEXTN_EMBED_TOKENS,
LLM_TENSOR_NEXTN_ENORM,
LLM_TENSOR_NEXTN_HNORM,
LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
};
enum llm_tensor_layer {

16
src/llama-context.cpp
View file

@ -786,7 +786,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;
const int64_t n_embd = hparams.n_embd;
const int32_t n_vocab = model.vocab.n_tokens();
const int64_t n_vocab = model.vocab.n_tokens();
// note: during encode, we always pass the full sequence starting from pos = 0
if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
@ -959,7 +959,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
const auto & vocab = model.vocab;
const auto & hparams = model.hparams;
const int32_t n_vocab = vocab.n_tokens();
const int64_t n_vocab = vocab.n_tokens();
const int64_t n_embd = hparams.n_embd;
// when computing embeddings, all tokens are output
@ -1328,21 +1328,21 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
}
void llama_context::output_reorder() {
const uint32_t n_vocab = model.vocab.n_tokens();
const uint64_t n_vocab = model.vocab.n_tokens();
const uint64_t n_embd = model.hparams.n_embd;
for (uint32_t s = 0; s < output_swaps.size(); ++s) {
const uint32_t i0 = output_swaps[s].i0;
const uint32_t i1 = output_swaps[s].i1;
for (size_t s = 0; s < output_swaps.size(); ++s) {
const uint64_t i0 = output_swaps[s].i0;
const uint64_t i1 = output_swaps[s].i1;
if (logits_size > 0) {
for (uint32_t k = 0; k < n_vocab; k++) {
for (uint64_t k = 0; k < n_vocab; k++) {
std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]);
}
}
if (embd_size > 0) {
for (uint32_t k = 0; k < n_embd; k++) {
for (uint64_t k = 0; k < n_embd; k++) {
std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
}
}

8
src/llama-graph.cpp
View file

@ -749,8 +749,8 @@ ggml_tensor * llm_graph_context::build_ffn(
if (down) {
cur = build_lora_mm(down, cur);
if (arch == LLM_ARCH_GLM4) {
// GLM4 seems to have numerical issues with half-precision accumulators
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
}
@ -1391,8 +1391,8 @@ ggml_tensor * llm_graph_context::build_attn(
if (wo) {
cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4) {
// GLM4 seems to have numerical issues with half-precision accumulators
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
}

1
src/llama-hparams.h
View file

@ -73,6 +73,7 @@ struct llama_hparams {
bool expert_weights_norm = false;
uint32_t expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
uint32_t moe_every_n_layers = 0;
uint32_t nextn_predict_layers = 0;
float f_norm_eps;
float f_norm_rms_eps;

4
src/llama-kv-cache-unified.cpp
View file

@ -39,6 +39,10 @@ llama_kv_cache_unified::llama_kv_cache_unified(
if (model.arch == LLM_ARCH_GEMMA3N) {
n_layer_cache = 20;
}
if (model.arch == LLM_ARCH_GLM4_MOE) {
// GLM-4.5: Only process up to last layer, skip final NextN layer
n_layer_cache = hparams.n_layer - hparams.nextn_predict_layers;
}
// create a context for each buffer type
std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;

5
src/llama-model-loader.h
View file

@ -58,8 +58,9 @@ struct llama_model_loader {
}
};
static const int TENSOR_NOT_REQUIRED = 1;
static const int TENSOR_DUPLICATED = 2;
static const int TENSOR_NOT_REQUIRED = 1 << 0;
static const int TENSOR_DUPLICATED = 1 << 1;
static const int TENSOR_SKIP = 1 << 2;
int n_kv = 0;
int n_tensors = 0;

300
src/llama-model.cpp
View file

@ -114,8 +114,10 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_A13B: return "A13B";
case LLM_TYPE_21B_A3B: return "21B.A3B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_106B_A12B: return "106B.A12B";
case LLM_TYPE_235B_A22B: return "235B.A22B";
case LLM_TYPE_300B_A47B: return "300B.A47B";
case LLM_TYPE_355B_A32B: return "355B.A32B";
case LLM_TYPE_E2B: return "E2B";
case LLM_TYPE_E4B: return "E4B";
default: return "?B";
@ -1439,6 +1441,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_GLM4_MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
// MoE parameters
ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert);
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
// Expert gating function (GLM-4.5 uses sigmoid)
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
}
// NextN/MTP parameters
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
switch (hparams.n_layer) {
case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_BITNET:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@ -1961,6 +1991,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const auto TENSOR_DUPLICATED = llama_model_loader::TENSOR_DUPLICATED;
const auto TENSOR_NOT_REQUIRED = llama_model_loader::TENSOR_NOT_REQUIRED;
const auto TENSOR_SKIP = llama_model_loader::TENSOR_SKIP;
// create tensors for the weights
{
@ -2020,7 +2051,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
// skip unused tensors
if (info.op == GGML_OP_NONE) {
if (info.op == GGML_OP_NONE || flags & TENSOR_SKIP) {
const size_t nbytes = ggml_nbytes(t_meta);
LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes);
@ -4523,6 +4554,105 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
}
} break;
case LLM_ARCH_GLM4_MOE:
{
const int64_t n_expert = hparams.n_expert;
const int64_t n_expert_used = hparams.n_expert_used;
const int64_t n_expert_shared = hparams.n_expert_shared;
GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
// 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 }, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
}
// Load ALL tensors including NextN layer to satisfy total tensor count
// but only PROCESS up to last layer (skipping final NextN layer) in forward pass
for (int i = 0; i < n_layer; ++i) {
int flags = 0;
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
// skip all tensors in the NextN layers
flags |= TENSOR_SKIP;
}
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
// GLM-style attention with bias terms
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.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
// K/Q norm tensors (optional for GLM-4.5 355B variant)
layer.attn_q_norm = create_tensor(
tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
layer.attn_k_norm = create_tensor(
tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags);
// Check if this layer uses MoE or dense FFN based on n_layer_dense_lead
// GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE
const bool use_moe = (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead);
if (use_moe) {
// MoE layers
layer.ffn_gate_inp =
create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags);
// MoE branch
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
layer.ffn_gate_exps = create_tensor(
tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
layer.ffn_down_exps = create_tensor(
tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags);
layer.ffn_up_exps = create_tensor(
tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
// Shared expert
if (n_expert_shared > 0) {
const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
layer.ffn_gate_shexp = create_tensor(
tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
layer.ffn_down_shexp = create_tensor(
tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
layer.ffn_up_shexp = create_tensor(
tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
}
} else {
// Dense layers (first k layers) - GLM uses separate gate/up projections
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, flags);
}
// NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
}
}
}
break;
case LLM_ARCH_NEMOTRON:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@ -13664,6 +13794,169 @@ struct llm_build_glm4 : public llm_graph_context {
}
};
struct llm_build_glm4_moe : public llm_graph_context {
llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv_unified();
ggml_tensor * inp_out_ids = build_inp_out_ids();
// Only process up to last layer (skip final NextN layer)
// Final layer tensors are loaded but not processed in forward pass
const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
for (int il = 0; il < n_transformer_layers; ++il) {
ggml_tensor * inpSA = inpL;
// Pre-attention norm
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self-attention
{
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
}
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
}
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
}
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// Apply Q/K norm if available (GLM-4.5 355B variant)
if (model.layers[il].attn_q_norm) {
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
}
if (model.layers[il].attn_k_norm) {
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il);
}
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_transformer_layers - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// Post-attention norm
cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "post_attn_norm", il);
// Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
// Dense FFN layer
cur = 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, "ffn_out", il);
} else {
// MoE layer with shared experts
const int64_t n_expert = hparams.n_expert;
const int64_t n_expert_used = hparams.n_expert_used;
// Process routed experts using existing MoE infrastructure
ggml_tensor * routed_out = 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, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(routed_out, "ffn_moe_out", il);
// Process shared expert on original input
ggml_tensor * shared_out = build_ffn(cur,
model.layers[il].ffn_up_shexp, NULL, NULL,
model.layers[il].ffn_gate_shexp, NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(shared_out, "ffn_shexp_out", il);
// Final output: routed_output + shared_output
cur = ggml_add(ctx0, routed_out, shared_out);
cb(cur, "ffn_out", il);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
};
struct llm_build_nemotron : public llm_graph_context {
llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@ -17977,6 +18270,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_glm4>(*this, params);
} break;
case LLM_ARCH_GLM4_MOE:
{
llm = std::make_unique<llm_build_glm4_moe>(*this, params);
} break;
case LLM_ARCH_BITNET:
{
llm = std::make_unique<llm_build_bitnet>(*this, params);
@ -18308,6 +18605,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_HUNYUAN_DENSE:
case LLM_ARCH_LFM2:
case LLM_ARCH_SMALLTHINKER:
case LLM_ARCH_GLM4_MOE:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:

13
src/llama-model.h
View file

@ -101,8 +101,10 @@ enum llm_type {
LLM_TYPE_A13B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_235B_A22B,
LLM_TYPE_300B_A47B, // Ernie MoE big
LLM_TYPE_355B_A32B, // GLM-4.5
LLM_TYPE_E2B,
LLM_TYPE_E4B,
};
@ -166,6 +168,15 @@ struct llama_layer_shortconv {
struct ggml_tensor * out_proj = nullptr;
};
struct llama_layer_nextn {
struct ggml_tensor * eh_proj = nullptr;
struct ggml_tensor * embed_tokens = nullptr;
struct ggml_tensor * enorm = nullptr;
struct ggml_tensor * hnorm = nullptr;
struct ggml_tensor * shared_head_head = nullptr;
struct ggml_tensor * shared_head_norm = nullptr;
};
struct llama_layer {
// normalization
struct ggml_tensor * attn_norm = nullptr;
@ -354,6 +365,8 @@ struct llama_layer {
struct llama_layer_convnext convnext;
struct llama_layer_shortconv shortconv;
struct llama_layer_nextn nextn;
};
struct llama_model {

3
src/llama-vocab.cpp
View file

@ -2430,6 +2430,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<｜fim▁begin｜>" // DeepSeek
|| t.first == "<PRE>"
|| t.first == "▁<PRE>" // CodeLlama
|| t.first == "<|code_prefix|>" // GLM-4.5
) {
special_fim_pre_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2449,6 +2450,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<｜fim▁hole｜>" // DeepSeek
|| t.first == "<SUF>"
|| t.first == "▁<SUF>" // CodeLlama
|| t.first == "<|code_suffix|>" // GLM-4.5
) {
special_fim_suf_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@ -2468,6 +2470,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<｜fim▁end｜>" // DeepSeek
|| t.first == "<MID>"
|| t.first == "▁<MID>" // CodeLlama
|| t.first == "<|code_middle|>" // GLM-4.5
) {
special_fim_mid_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {