Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '3ca19b0e9f3f4f444d22c9f509805d037a611847' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	common/CMakeLists.txt
#	common/chat-peg-parser.cpp
#	docs/backend/SYCL.md
#	docs/ops.md
#	docs/ops/SYCL.csv
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/convert.hpp
#	ggml/src/ggml-sycl/element_wise.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/norm.cpp
#	ggml/src/ggml-sycl/rope.cpp
#	ggml/src/ggml-sycl/rope.hpp
#	ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl
#	scripts/compare-llama-bench.py
#	scripts/sync_vendor.py
#	tests/CMakeLists.txt
#	tools/cli/cli.cpp
This commit is contained in:
Concedo
2026-03-15 11:11:31 +08:00
parent 3b9385a627 3ca19b0e9f
commit 67c9798d0b
41 changed files with 836 additions and 212 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+9
View File
@@ -185,6 +185,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EXPERT_GROUP_SCALE, "%s.expert_group_scale" },
{ LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" },
{ LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
{ LLM_KV_MOE_LATENT_SIZE, "%s.moe_latent_size" },
{ LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
{ LLM_KV_NUM_DEEPSTACK_LAYERS, "%s.n_deepstack_layers" },
{ LLM_KV_POOLING_TYPE, "%s.pooling_type" },
@@ -365,6 +366,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
{ LLM_TENSOR_FFN_LATENT_DOWN, "blk.%d.ffn_latent_down" },
{ LLM_TENSOR_FFN_LATENT_UP, "blk.%d.ffn_latent_up" },
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
@@ -1087,6 +1090,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
@@ -1878,6 +1882,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_FFN_LATENT_DOWN,
LLM_TENSOR_FFN_LATENT_UP,
// MoE shared expert layer
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
@@ -2753,6 +2759,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{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}},
// Nemotron 3 Super
{LLM_TENSOR_FFN_LATENT_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_LATENT_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
};
LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
src/llama-arch.h
+3
View File
@@ -189,6 +189,7 @@ enum llm_kv {
LLM_KV_EXPERT_GROUP_SCALE,
LLM_KV_EXPERTS_PER_GROUP,
LLM_KV_MOE_EVERY_N_LAYERS,
LLM_KV_MOE_LATENT_SIZE,
LLM_KV_NEXTN_PREDICT_LAYERS,
LLM_KV_NUM_DEEPSTACK_LAYERS,
LLM_KV_POOLING_TYPE,
@@ -385,6 +386,8 @@ enum llm_tensor {
LLM_TENSOR_FFN_GATE_CHEXPS,
LLM_TENSOR_FFN_UP_CHEXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
LLM_TENSOR_FFN_LATENT_DOWN,
LLM_TENSOR_FFN_LATENT_UP,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
src/llama-graph.cpp
+2 -2
View File
@@ -250,7 +250,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
const bool last = (
cparams.pooling_type == LLAMA_POOLING_TYPE_LAST ||
(cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && arch == LLM_ARCH_QWEN3) // qwen3 reranking & embedding models use last token
(cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && (arch == LLM_ARCH_QWEN3 || arch == LLM_ARCH_QWEN3VL)) // qwen3 reranking & embedding models use last token
);
for (int i = 0; i < n_tokens; ++i) {
@@ -2552,7 +2552,7 @@ void llm_graph_context::build_pooling(
}
// softmax for qwen3 reranker
if (arch == LLM_ARCH_QWEN3) {
if (arch == LLM_ARCH_QWEN3 || arch == LLM_ARCH_QWEN3VL) {
cur = ggml_soft_max(ctx0, cur);
}
} break;
src/llama-hparams.h
+1
View File
@@ -89,6 +89,7 @@ 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 moe_latent_size = 0;
uint32_t nextn_predict_layers = 0;
float f_norm_eps;
src/llama-model.cpp
+9 -2
View File
@@ -249,6 +249,7 @@ const char * llm_type_name(llm_type type) {
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_120B_A12B: return "120B.A12B";
case LLM_TYPE_122B_A10B: return "122B.A10B";
case LLM_TYPE_196B_A11B: return "196B.A11B";
case LLM_TYPE_230B_A10B: return "230B.A10B";
@@ -1975,10 +1976,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared, false);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
ml.get_key(LLM_KV_MOE_LATENT_SIZE, hparams.moe_latent_size, false);
switch (hparams.n_layer) {
case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
case 56: type = LLM_TYPE_9B; break;
case 88: type = LLM_TYPE_120B_A12B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
@@ -5702,6 +5705,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t n_ssm_head = hparams.ssm_dt_rank;
const int64_t n_group = hparams.ssm_n_group;
const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_ssm_head;
const int64_t moe_n_embd = hparams.moe_latent_size > 0 ? hparams.moe_latent_size : n_embd;
// embeddings
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5761,8 +5765,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert }, 0);
// MoE branch
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);
layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_latent_up = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP, "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, moe_n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {moe_n_embd, n_ff_exp, n_expert}, 0);
// Shared expert branch
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
src/llama-model.h
+5
View File
@@ -126,6 +126,7 @@ enum llm_type {
LLM_TYPE_100B_A6B,
LLM_TYPE_102B_A12B, // Solar-Open
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_120B_A12B, // Nemotron 3 Super
LLM_TYPE_122B_A10B, // Qwen3.5
LLM_TYPE_196B_A11B, // Step3.5-Flash
LLM_TYPE_230B_A10B, // Minimax M2
@@ -294,6 +295,10 @@ struct llama_layer {
struct ggml_tensor * ffn_up_exps_b = nullptr;
struct ggml_tensor * ffn_gate_up_exps_b = nullptr;
// ff MoE latent proj
struct ggml_tensor * ffn_latent_down = nullptr;
struct ggml_tensor * ffn_latent_up = nullptr;
// ff shared expert (shexp)
struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
struct ggml_tensor * ffn_gate_shexp = nullptr;
src/llama-quant.cpp
+16 -13
View File
@@ -872,9 +872,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
quantize_state_impl qs(model, params);
// these need to be set to n_layer by default
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
if (params->only_copy) {
ftype = ml.ftype;
}
@@ -981,6 +978,22 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
// compute tensor metadata once and cache it
std::vector<tensor_metadata> metadata(tensors.size());
// initialize quantization state before preliminary loop (counters for use_more_bits)
{
for (size_t i = 0; i < tensors.size(); ++i) {
const auto cat = tensor_get_category(tensors[i]->tensor->name);
if (category_is_attn_v(cat)) {
++qs.n_attention_wv;
}
if (cat == tensor_category::OUTPUT) {
qs.has_tied_embeddings = false;
}
metadata[i].category = cat; // save and re-use the category while we're at it
}
// these also need to be set to n_layer by default
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer;
}
// flag for --dry-run
bool will_require_imatrix = false;
@@ -993,16 +1006,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
const struct ggml_tensor * tensor = it->tensor;
const std::string name = ggml_get_name(tensor);
metadata[i].category = tensor_get_category(name);
if (category_is_attn_v(metadata[i].category)) {
++qs.n_attention_wv;
}
if (tensor_name_match_output_weight(name.c_str())) {
qs.has_tied_embeddings = false;
}
uint16_t i_split = params->keep_split ? it->idx : 0;
if (!ctx_outs[i_split]) {
ctx_outs[i_split].reset(gguf_init_empty());
src/models/nemotron-h.cpp
+18 -4
View File
@@ -114,9 +114,18 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
ggml_tensor * ffn_inp = cur;
ggml_tensor * inp_emb = cur;
ggml_tensor * inp_latent = cur;
if (model.layers[il].ffn_latent_down) {
inp_latent = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_down, cur);
}
ggml_tensor * router_logits = build_lora_mm(model.layers[il].ffn_gate_inp, cur);
cb(router_logits, "ffn_moe_logits", il);
ggml_tensor * moe_out =
build_moe_ffn(ffn_inp,
build_moe_ffn(inp_latent,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
nullptr, // no gate
@@ -126,10 +135,15 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
il);
il,
router_logits);
cb(moe_out, "ffn_moe_out", il);
ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
if (model.layers[il].ffn_latent_up) {
moe_out = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_up, moe_out);
}
ggml_tensor * ffn_shexp = build_ffn(inp_emb,
model.layers[il].ffn_up_shexp, NULL, NULL,
NULL /* no gate */ , NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,