vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-11 21:32:11 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

Merge commit 'ddf9f94389' into concedo_experimental

Browse source 
# Conflicts:
#	examples/model-conversion/scripts/causal/run-converted-model.sh
#	examples/model-conversion/scripts/causal/run-org-model.py
#	src/CMakeLists.txt
#	src/llama-quant.cpp
#	tools/server/README.md
This commit is contained in:
Concedo 2025-11-28 23:27:50 +08:00
commit 0ccb298087
23 changed files with 2813 additions and 86 deletions
Download patch file Download diff file Expand all files Collapse all files

35
src/llama-arch.cpp
View file

@ -32,6 +32,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_QWEN2VL, "qwen2vl" },
{ LLM_ARCH_QWEN3, "qwen3" },
{ LLM_ARCH_QWEN3MOE, "qwen3moe" },
{ LLM_ARCH_QWEN3NEXT, "qwen3next" },
{ LLM_ARCH_QWEN3VL, "qwen3vl" },
{ LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" },
{ LLM_ARCH_PHI2, "phi2" },
@ -829,6 +830,38 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_QWEN3NEXT,
{
{ 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_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_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_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
{ 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_SSM_A, "blk.%d.ssm_a" },
{ LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
{ LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
{ LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" },
{ LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
{ LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
{ LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
},
},
{
LLM_ARCH_QWEN3VL,
{
@ -2556,6 +2589,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@ -2754,6 +2788,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_NEMOTRON_H:
case LLM_ARCH_QWEN3NEXT:
return true;
default:
return false;

2
src/llama-arch.h
View file

@ -36,6 +36,7 @@ enum llm_arch {
LLM_ARCH_QWEN2VL,
LLM_ARCH_QWEN3,
LLM_ARCH_QWEN3MOE,
LLM_ARCH_QWEN3NEXT,
LLM_ARCH_QWEN3VL,
LLM_ARCH_QWEN3VLMOE,
LLM_ARCH_PHI2,
@ -381,6 +382,7 @@ enum llm_tensor {
LLM_TENSOR_SSM_D,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next
LLM_TENSOR_TIME_MIX_W0,
LLM_TENSOR_TIME_MIX_W1,
LLM_TENSOR_TIME_MIX_W2,

4
src/llama-context.cpp
View file

@ -1,5 +1,6 @@
#include "llama-context.h"
#include "llama-arch.h"
#include "llama-impl.h"
#include "llama-batch.h"
#include "llama-io.h"
@ -1388,6 +1389,9 @@ void llama_context::output_reorder() {
//
uint32_t llama_context::graph_max_nodes() const {
if (model.arch == LLM_ARCH_QWEN3NEXT) {
return std::max<uint32_t>(8192u, 32u*model.n_tensors());
}
return std::max<uint32_t>(1024u, 8u*model.n_tensors());
}

2
src/llama-hparams.h
View file

@ -6,7 +6,7 @@
// bump if necessary
#define LLAMA_MAX_LAYERS 512
#define LLAMA_MAX_EXPERTS 384 // Kimi-K2
#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
enum llama_expert_gating_func_type {
LLAMA_EXPERT_GATING_FUNC_TYPE_NONE = 0,

101
src/llama-model.cpp
View file

@ -2,7 +2,6 @@
#include "llama-impl.h"
#include "llama-mmap.h"
#include "llama-batch.h"
#include "llama-cparams.h"
#include "llama-model-loader.h"
@ -106,6 +105,7 @@
#include "models/qwen3vl.cpp"
#include "models/qwen3vl-moe.cpp"
#include "models/qwen3moe.cpp"
#include "models/qwen3next.cpp"
#include "models/refact.cpp"
#include "models/rnd1.cpp"
#include "models/rwkv6-base.cpp"
@ -2328,6 +2328,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN3NEXT:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
// Load linear attention (gated delta net) parameters
ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv);
ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner);
ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval"
}
switch (hparams.n_layer) {
case 80: type = LLM_TYPE_80B_A3B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
default: throw std::runtime_error("unsupported model architecture");
}
@ -6571,6 +6594,74 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
} break;
case LLM_ARCH_QWEN3NEXT:
{
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);
}
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
// Calculate dimensions from hyperparameters
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t head_v_dim = hparams.ssm_d_state;
const int64_t n_k_heads = hparams.ssm_n_group;
const int64_t n_v_heads = hparams.ssm_dt_rank;
const int64_t key_dim = head_k_dim * n_k_heads;
const int64_t value_dim = head_v_dim * n_v_heads;
const int64_t conv_dim = key_dim * 2 + value_dim;
// Calculate projection sizes
const int64_t qkvz_dim = key_dim * 2 + value_dim * 2;
const int64_t ba_dim = n_v_heads * 2;
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.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
if (!hparams.is_recurrent(i)) {
// Attention layers
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// Q/K normalization for attention layers
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);
} else {
// Linear attention (gated delta net) specific tensors
// Create tensors with calculated dimensions
layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), { n_embd, qkvz_dim }, 0);
layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0);
layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), { hparams.ssm_dt_rank }, 0);
layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0);
layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0);
}
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_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);
// Shared experts
layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, hparams.n_ff_shexp }, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, hparams.n_ff_shexp }, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -6842,6 +6933,7 @@ void llama_model::print_info() const {
arch == LLM_ARCH_FALCON_H1 ||
arch == LLM_ARCH_PLAMO2 ||
arch == LLM_ARCH_GRANITE_HYBRID ||
arch == LLM_ARCH_QWEN3NEXT ||
arch == LLM_ARCH_NEMOTRON_H) {
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
@ -7586,7 +7678,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
case LLM_ARCH_PANGU_EMBED:
{
llm = std::make_unique<llm_build_pangu_embedded>(*this, params);
}break;
} break;
case LLM_ARCH_QWEN3NEXT:
{
llm = std::make_unique<llm_build_qwen3next>(*this, params);
} break;
default:
GGML_ABORT("fatal error");
}
@ -7813,6 +7909,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_COGVLM:
case LLM_ARCH_PANGU_EMBED:
case LLM_ARCH_AFMOE:
case LLM_ARCH_QWEN3NEXT:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:

4
src/llama-model.h
View file

@ -113,6 +113,7 @@ enum llm_type {
LLM_TYPE_16B_A1B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_80B_A3B, // Qwen3 Next
LLM_TYPE_100B_A6B,
LLM_TYPE_106B_A12B, // GLM-4.5-Air
LLM_TYPE_230B_A10B, // Minimax M2
@ -309,6 +310,9 @@ struct llama_layer {
struct ggml_tensor * ssm_conv1d_b = nullptr;
struct ggml_tensor * ssm_dt_b = nullptr;
// qwen3next
struct ggml_tensor * ssm_beta_alpha = nullptr;
// rwkv
struct ggml_tensor * time_mix_w1 = nullptr;
struct ggml_tensor * time_mix_w2 = nullptr;

18
src/llama-quant.cpp
View file

@ -684,7 +684,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
}
LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
continue;
} else if (remapped_name != it.first) {
}
if (remapped_name != it.first) {
ggml_set_name(it.second.tensor, remapped_name.c_str());
LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
}
@ -729,13 +731,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
{
const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
// attention layers have a non-zero number of kv heads
int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
if (llama_model_has_encoder(&model)) {
// now n_attn_layer is the number of attention layers in the encoder
// now n_layer_attn is the number of attention layers in the encoder
// for each decoder block, there are 2 attention layers
n_attn_layer += 2 * model.hparams.dec_n_layer;
n_layer_attn += 2 * model.hparams.dec_n_layer;
}
GGML_ASSERT_CONTINUE((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
// note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
GGML_ASSERT_CONTINUE((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
}
size_t total_size_org = 0;

52
src/models/models.h
View file

@ -2,8 +2,9 @@
#include "../llama-model.h"
#include "../llama-graph.h"
#include "../llama-memory-recurrent.h"
// TODO: remove in follow-up PR - move to .cpp files
#include "../llama-memory-recurrent.h"
#include <cmath>
struct llm_graph_context_mamba : public llm_graph_context {
@ -421,7 +422,56 @@ struct llm_build_qwen3vl : public llm_graph_context {
struct llm_build_qwen3vlmoe : public llm_graph_context {
llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_qwen3next : public llm_graph_context_mamba {
llm_build_qwen3next(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
llm_graph_input_attn_kv * inp_attn,
ggml_tensor * cur,
ggml_tensor * inp_pos,
int il);
ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
int il);
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
int il);
ggml_tensor * build_delta_net_recurrent(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
int il);
ggml_tensor * build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
int layer);
const llama_model & model;
};
struct llm_build_qwen : public llm_graph_context {
llm_build_qwen(const llama_model & model, const llm_graph_params & params);

1042
src/models/qwen3next.cpp Normal file
View file

File diff suppressed because it is too large Load diff