Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-10 14:59:11 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '8f8f2274ee3601fecf6e2d57b52f701c81bede21' into concedo_experimental

Browse Source 
# Conflicts:
#	.devops/rocm.Dockerfile
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	CMakeLists.txt
#	examples/simple/simple.cpp
#	ggml/src/ggml-cann/common.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-opencl/kernels/tsembd.cl
#	ggml/src/ggml-sycl/binbcast.cpp
#	ggml/src/ggml-sycl/binbcast.hpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/tsembd.cpp
#	ggml/src/ggml-zdnn/ggml-zdnn.cpp
#	src/llama-model.cpp
#	tools/batched-bench/CMakeLists.txt
#	tools/cvector-generator/CMakeLists.txt
#	tools/export-lora/CMakeLists.txt
#	tools/gguf-split/CMakeLists.txt
#	tools/imatrix/CMakeLists.txt
#	tools/llama-bench/CMakeLists.txt
#	tools/llama-bench/llama-bench.cpp
#	tools/main/CMakeLists.txt
#	tools/main/README.md
#	tools/mtmd/CMakeLists.txt
#	tools/perplexity/CMakeLists.txt
#	tools/perplexity/perplexity.cpp
#	tools/quantize/CMakeLists.txt
#	tools/rpc/rpc-server.cpp
#	tools/run/CMakeLists.txt
#	tools/run/run.cpp
#	tools/tokenize/CMakeLists.txt
#	tools/tts/CMakeLists.txt
This commit is contained in:
Concedo
2025-09-21 08:58:23 +08:00
parent 04498a345a 8f8f2274ee
commit 3e72aaff5b
41 changed files with 1998 additions and 1463 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+327 -81
View File
@@ -41,6 +41,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_80M: return "80M";
case LLM_TYPE_109M: return "109M";
case LLM_TYPE_137M: return "137M";
case LLM_TYPE_140M: return "140M";
case LLM_TYPE_160M: return "160M";
case LLM_TYPE_190M: return "190M";
case LLM_TYPE_220M: return "220M";
@@ -49,6 +50,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_270M: return "270M";
case LLM_TYPE_335M: return "335M";
case LLM_TYPE_350M: return "350M";
case LLM_TYPE_360M: return "360M";
case LLM_TYPE_410M: return "410M";
case LLM_TYPE_450M: return "450M";
case LLM_TYPE_475M: return "475M";
@@ -56,6 +58,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_700M: return "700M";
case LLM_TYPE_770M: return "770M";
case LLM_TYPE_780M: return "780M";
case LLM_TYPE_950M: return "950M";
case LLM_TYPE_0_3B: return "0.3B";
case LLM_TYPE_0_5B: return "0.5B";
case LLM_TYPE_0_6B: return "0.6B";
@@ -627,19 +630,32 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP, hparams.n_moe_layer_step);
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.set_swa_pattern(4); // pattern: 3 chunked - 1 full
const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
if (found_swa && hparams.n_swa == 0) {
hparams.swa_type = LLAMA_SWA_TYPE_NONE;
hparams.n_no_rope_layer_step = hparams.n_layer; // always use rope
} else {
hparams.swa_type = LLAMA_SWA_TYPE_CHUNKED;
hparams.n_swa = 8192;
hparams.set_swa_pattern(4); // pattern: 3 chunked - 1 full
}
switch (hparams.n_expert) {
case 0: {
// MobileLLM (no MoE)
switch (hparams.n_embd) {
case 2048: type = LLM_TYPE_140M; break;
case 4096: type = LLM_TYPE_360M; break;
case 6144: type = LLM_TYPE_950M; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case 16: type = LLM_TYPE_17B_16E; break;
case 128: type = LLM_TYPE_17B_128E; break;
default: type = LLM_TYPE_UNKNOWN;
}
if (type == LLM_TYPE_17B_128E) {
hparams.use_kq_norm = false;
}
hparams.use_kq_norm = type != LLM_TYPE_17B_128E;
} break;
case LLM_ARCH_ARCEE:
{
@@ -690,7 +706,30 @@ void llama_model::load_hparams(llama_model_loader & ml) {
} break;
case LLM_ARCH_GROK:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
// defaults for old GGUFs
hparams.yarn_beta_fast = 8.0f;
hparams.f_logit_scale = 0.5773502691896257f;
hparams.f_embedding_scale = 78.38367176906169f;
hparams.f_attn_out_scale = 0.08838834764831845f;
hparams.f_attn_logit_softcapping = 30.0f;
hparams.f_router_logit_softcapping = 30.0f;
// no final_logit_softcapping in grok-1
hparams.f_final_logit_softcapping = 0.0f;
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale, false);
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false);
ml.get_key(LLM_KV_ATTENTION_OUTPUT_SCALE, hparams.f_attn_out_scale, false);
ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false);
ml.get_key(LLM_KV_ROUTER_LOGIT_SOFTCAPPING, hparams.f_router_logit_softcapping, false);
ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.attn_temp_length, false);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, hparams.yarn_ext_factor, false);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, hparams.yarn_attn_factor, false);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast, false);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow, false);
switch (hparams.n_layer) {
case 64: type = LLM_TYPE_314B; break;
@@ -918,6 +957,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.causal_attn = false;
}
break;
case LLM_ARCH_LLADA_MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
// diffusion language model uses non-causal attention
hparams.causal_attn = false;
switch (hparams.n_layer) {
case 16: type = LLM_TYPE_A1_7B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN2MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
@@ -1320,6 +1371,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
if (found_swa && hparams.n_swa > 0) {
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
hparams.set_swa_pattern(4);
} else {
hparams.swa_type = LLAMA_SWA_TYPE_NONE;
}
switch (hparams.n_layer) {
case 16: type = LLM_TYPE_1B; break;
case 32: type = LLM_TYPE_7B; break;
@@ -2422,6 +2481,40 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
}
break;
case LLM_ARCH_LLADA_MOE:
{
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}, 0);
GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for llada-moe");
GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for llada-moe");
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
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}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0);
}
} break;
case LLM_ARCH_LLAMA4:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -2435,9 +2528,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Llama 4 requires n_moe_layer_step > 0");
for (int i = 0; i < n_layer; ++i) {
bool is_moe_layer = (i + 1) % hparams.n_moe_layer_step == 0;
bool is_moe_layer = hparams.n_moe_layer_step > 0 && (i + 1) % hparams.n_moe_layer_step == 0;
auto & layer = layers[i];
@@ -2598,6 +2690,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff/* / n_expert_used*/; // grok-1 n_ff_exp == n_ff
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
@@ -2612,50 +2705,19 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
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}, TENSOR_NOT_REQUIRED);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
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, n_ff, n_expert}, TENSOR_NOT_REQUIRED);
if (layer.ffn_gate_exps) {
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
} else {
// merge split expert into a single tensor for compatibility with older MIXTRAL models
// requires disabling mmap
//slaren removed this in #10026, but i think its useful to keep.
use_mmap_buffer = false;
ml.use_mmap = false;
if(!old_mixtral_warning_showed)
{
std::cout << "\n!!!!!!\nWARNING: Using extremely outdated MoE quant. Please update it!\nAttempting to apply hacky kcpp fallback, using last ctx:" << last_used_ctx << "\n";
old_mixtral_warning_showed = true;
}
ggml_context * ctx_split = last_used_ctx;
// for(auto it = ctx_map.cbegin(); it != ctx_map.cend(); ++it)
// {
// std::cout << "\nName: " << ggml_backend_buft_name(it->first) << " Addr: " << it->second << "\n";
// ctx_split = it->second;
// }
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
ggml_type type_gate = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).str().c_str())->type;
ggml_type type_down = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).str().c_str())->type;
ggml_type type_up = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, 0).str().c_str())->type;
layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd, n_ff, n_expert);
layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down, n_ff, n_embd, n_expert);
layer.ffn_up_exps = ggml_new_tensor_3d(ctx_split, type_up, n_embd, n_ff, n_expert);
ggml_set_name(layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i).str().c_str());
ggml_set_name(layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i).str().c_str());
ggml_set_name(layer.ffn_up_exps, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i).str().c_str());
for (uint32_t x = 0; x < n_expert; ++x) {
// the individual experts are loaded into a view of the merged tensor
ml.create_tensor_as_view(ctx_split, layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_gate_exps->nb[2]*x);
ml.create_tensor_as_view(ctx_split, layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd }, layer.ffn_down_exps->nb[2]*x);
ml.create_tensor_as_view(ctx_split, layer.ffn_up_exps, tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_up_exps->nb[2]*x);
}
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
if (!layer.ffn_post_norm) {
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
}
layer.layer_out_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
}
} break;
case LLM_ARCH_DBRX:
@@ -6343,6 +6405,14 @@ struct llm_build_llama : public llm_graph_context {
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
if (hparams.use_kq_norm) {
// Llama4TextL2Norm
Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
cb(Qcur, "Qcur_normed", il);
cb(Kcur, "Kcur_normed", il);
}
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
@@ -6450,7 +6520,8 @@ struct llm_build_llama_iswa : public llm_graph_context {
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
(il + 1) % hparams.n_no_rope_layer_step != 0;
// norm
cur = build_norm(inpL,
@@ -7128,9 +7199,6 @@ struct llm_build_grok : public llm_graph_context {
inpL = build_inp_embd(model.tok_embd);
// multiply by embedding_multiplier_scale of 78.38367176906169
inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -7202,26 +7270,22 @@ struct llm_build_grok : public llm_graph_context {
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// Grok
// if attn_out_norm is present then apply it before adding the input
if (model.layers[il].attn_out_norm) {
cur = build_norm(cur,
model.layers[il].attn_out_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_out_norm", il);
}
cur = build_norm(cur,
model.layers[il].attn_out_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_out_norm", il);
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network
// MoE branch
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
cur = build_moe_ffn(cur,
// MoE branch
ggml_tensor * moe_out = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
@@ -7232,18 +7296,28 @@ struct llm_build_grok : public llm_graph_context {
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur, "ffn_moe_out", il);
cb(moe_out, "ffn_moe_out", il);
// Grok
// if layer_out_norm is present then apply it before adding the input
// Idea: maybe ffn_out_norm is a better name
if (model.layers[il].layer_out_norm) {
cur = build_norm(cur,
model.layers[il].layer_out_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "layer_out_norm", il);
if (model.layers[il].ffn_up) {
ggml_tensor * ffn_out = 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_GELU, LLM_FFN_PAR, il);
cb(ffn_out, "ffn_out", il);
cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
cb(cur, "ffn_out", il);
} else {
cur = moe_out;
}
cur = build_norm(cur,
model.layers[il].ffn_post_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_post_norm", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
@@ -7266,10 +7340,14 @@ struct llm_build_grok : public llm_graph_context {
// lm_head
cur = build_lora_mm(model.output, cur);
// Grok
// multiply logits by output_multiplier_scale of 0.5773502691896257
cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
cur = ggml_scale(ctx0, cur, 0.5773502691896257f);
// final logit soft-capping
if (hparams.f_final_logit_softcapping) {
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
cur = ggml_tanh(ctx0, cur);
cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
}
cb(cur, "result_output", -1);
res->t_logits = cur;
@@ -12249,6 +12327,7 @@ struct llm_build_olmo : public llm_graph_context {
}
};
template <bool iswa>
struct llm_build_olmo2 : public llm_graph_context {
llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -12264,7 +12343,14 @@ struct llm_build_olmo2 : public llm_graph_context {
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv();
using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
inp_attn_type * inp_attn = nullptr;
if constexpr (iswa) {
inp_attn = build_attn_inp_kv_iswa();
} else {
inp_attn = build_attn_inp_kv();
}
ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12297,17 +12383,36 @@ struct llm_build_olmo2 : public llm_graph_context {
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);
Qcur = ggml_rope_ext(
const bool is_swa = hparams.is_swa(il);
if (is_swa) {
// For sliding window layers, Olmo3 use regular rope with no yarn rope scaling.
// This is achieved here by setting freq_scale and attn_factor to 1.
// We also set ext_factor to 0 to avoid a few unnecessary computations.
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
0.0, 1.0, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
0.0, 1.0, beta_fast, beta_slow
);
} else {
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(
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);
@@ -12506,6 +12611,132 @@ struct llm_build_olmoe : public llm_graph_context {
}
};
struct llm_build_llada_moe : public llm_graph_context {
llm_build_llada_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_ASSERT(n_embd_head == hparams.n_rot);
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_no_cache();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
// norm
cur = build_norm(inpL,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self_attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
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);
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
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, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 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);
// MoE branch
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
cur = 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,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(cur, "ffn_moe_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_openelm : public llm_graph_context {
llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -18698,6 +18929,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
//case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
case LLM_ARCH_DREAM:
case LLM_ARCH_LLADA:
case LLM_ARCH_LLADA_MOE:
{
res = nullptr;
} break;
@@ -18835,7 +19067,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
} break;
case LLM_ARCH_LLAMA4:
{
llm = std::make_unique<llm_build_llama_iswa>(*this, params);
if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
llm = std::make_unique<llm_build_llama>(*this, params);
} else {
llm = std::make_unique<llm_build_llama_iswa>(*this, params);
}
} break;
case LLM_ARCH_DECI:
{
@@ -18903,6 +19139,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
llm = std::make_unique<llm_build_llada>(*this, params);
}
break;
case LLM_ARCH_LLADA_MOE:
{
llm = std::make_unique<llm_build_llada_moe>(*this, params);
}
break;
case LLM_ARCH_QWEN2VL:
{
llm = std::make_unique<llm_build_qwen2vl>(*this, params);
@@ -19015,7 +19256,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
} break;
case LLM_ARCH_OLMO2:
{
llm = std::make_unique<llm_build_olmo2>(*this, params);
if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
llm = std::make_unique<llm_build_olmo2<true>>(*this, params);
} else {
llm = std::make_unique<llm_build_olmo2<false>>(*this, params);
}
} break;
case LLM_ARCH_OLMOE:
{
@@ -19369,6 +19614,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_QWEN2MOE:
case LLM_ARCH_QWEN3:
case LLM_ARCH_QWEN3MOE:
case LLM_ARCH_LLADA_MOE:
case LLM_ARCH_OLMO2:
case LLM_ARCH_OLMOE:
case LLM_ARCH_PHI2: