upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-03 08:30:48 -05:00
Code Issues Packages Projects Releases Wiki Activity

llama : load MTP tensors only if they are really used (#26296)

Browse Source 
* llama : load MTP tensors only if they are really used

* llama : skip loading MTP (if not used) in remaining models that support MTP

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming
2026-07-31 14:57:02 +02:00
committed by GitHub
parent 6f3c0a790b
commit 82dbc4f017
13 changed files with 65 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
common/common.cpp
View File

@@ -1620,6 +1620,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
mparams.progress_callback = params.load_progress_callback;
mparams.progress_callback_user_data = params.load_progress_callback_user_data;
mparams.no_alloc = params.no_alloc;
mparams.load_mtp = std::find(params.speculative.types.begin(), params.speculative.types.end(), COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
return mparams;
}

1
include/llama.h
View File

@@ -337,6 +337,7 @@ extern "C" {
bool use_extra_bufts; // use extra buffer types (used for weight repacking)
bool no_host; // bypass host buffer allowing extra buffers to be used
bool no_alloc; // only load metadata and simulate memory allocations
bool load_mtp; // whether to load MTP layers
};
struct llama_sampler_seq_config {

2
src/llama-model-loader.cpp
View File

@@ -526,6 +526,7 @@ llama_model_loader::llama_model_loader(
llama_load_mode load_mode,
bool check_tensors,
bool no_alloc,
bool load_mtp,
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p)
: metadata(meta), set_tensor_data(set_tensor_data), set_tensor_data_ud(set_tensor_data_ud) {
@@ -812,6 +813,7 @@ llama_model_loader::llama_model_loader(
this->check_tensors = check_tensors;
this->no_alloc = no_alloc;
this->load_mtp = load_mtp;
}
std::string llama_model_loader::get_arch_name() const {

2
src/llama-model-loader.h
View File

@@ -79,6 +79,7 @@ struct llama_model_loader {
bool use_direct_io = false;
bool check_tensors;
bool no_alloc;
bool load_mtp;
llama_files files;
llama_ftype ftype;
@@ -129,6 +130,7 @@ struct llama_model_loader {
llama_load_mode load_mode,
bool check_tensors,
bool no_alloc,
bool load_mtp,
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);

1
src/llama-model.cpp
View File

@@ -2390,6 +2390,7 @@ llama_model_params llama_model_default_params() {
/*.use_extra_bufts =*/ true,
/*.no_host =*/ false,
/*.no_alloc =*/ false,
/*.load_mtp =*/ false,
};
return result;

2
src/llama-quant.cpp
View File

@@ -893,7 +893,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
const llama_model_kv_override * kv_overrides = params->kv_overrides;
std::vector<std::string> splits = {};
llama_model_loader ml(/*metadata*/ nullptr, /*set_tensor_data*/ nullptr, /*set_tensor_data_ud*/ nullptr,
fname_inp, splits, /*file*/ nullptr, /*load_mode*/ load_mode, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
fname_inp, splits, /*file*/ nullptr, /*load_mode*/ load_mode, /*check_tensors*/ true, /*no_alloc*/ false, /*load_mtp*/ true, kv_overrides, nullptr);
ml.init_mappings(false); // no prefetching
auto mparams = llama_model_default_params();

2
src/llama.cpp
View File

@@ -305,7 +305,7 @@ static std::pair<int, llama_model *> llama_model_load(struct gguf_context * meta
const std::string & fname, std::vector<std::string> & splits, FILE * file, llama_model_params & params) {
try {
llama_model_loader ml(metadata, set_tensor_data, set_tensor_data_ud, fname, splits, file, params.load_mode,
params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
params.check_tensors, params.no_alloc, params.load_mtp, params.kv_overrides, params.tensor_buft_overrides);
ml.print_info();
std::unique_ptr<llama_model> model_ptr(llama_model_create(ml, params));

6
src/models/cohere2moe.cpp
View File

@@ -55,7 +55,11 @@ void llama_model_cohere2moe::load_arch_tensors(llama_model_loader & ml) {
const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
if (!ml.load_mtp) {
mtp_flags |= TENSOR_SKIP;
}
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);

6
src/models/glm-dsa.cpp
View File

@@ -91,7 +91,11 @@ void llama_model_glm_dsa::load_arch_tensors(llama_model_loader & ml) {
const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
if (!ml.load_mtp) {
mtp_flags |= TENSOR_SKIP;
}
const bool is_mla = hparams.is_mla();
if (!is_mla) {

6
src/models/hy-v3.cpp
View File

@@ -33,7 +33,11 @@ void llama_model_hy_v3::load_arch_tensors(llama_model_loader & ml) {
const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
if (!ml.load_mtp) {
mtp_flags |= TENSOR_SKIP;
}
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

31
src/models/qwen35.cpp
View File

@@ -39,6 +39,7 @@ void llama_model_qwen35::load_arch_tensors(llama_model_loader & ml) {
const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
@@ -97,25 +98,25 @@ void llama_model_qwen35::load_arch_tensors(llama_model_loader & ml) {
auto & layer = layers[il];
// MTP block looks like a full-attention Qwen3.5 decoder block.
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, 0);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, mtp_flags);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, mtp_flags);
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, 0);
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, mtp_flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, mtp_flags);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, mtp_flags);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, mtp_flags);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", il), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", il), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", il), {n_embd, n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", il), {n_embd, n_ff}, mtp_flags);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", il), { n_ff, n_embd}, mtp_flags);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", il), {n_embd, n_ff}, mtp_flags);
// NextN-specific tensors that define the MTP block.
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, 0);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, 0);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, 0);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, TENSOR_NOT_REQUIRED);
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, mtp_flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, mtp_flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, mtp_flags);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, mtp_flags|TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab }, mtp_flags|TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, mtp_flags|TENSOR_NOT_REQUIRED);
};
for (int i = 0; i < n_layer; ++i) {

39
src/models/qwen35moe.cpp
View File

@@ -42,6 +42,7 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
@@ -113,32 +114,32 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
// MTP block looks like a full-attention Qwen3.5 decoder block with MoE FFN.
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, 0);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, mtp_flags);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, mtp_flags);
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, 0);
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, mtp_flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, mtp_flags);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, mtp_flags);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, mtp_flags);
// Routed experts
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, 0);
create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, 0);
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, mtp_flags);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, mtp_flags);
create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, mtp_flags);
// Shared experts
layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, 0);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, n_ff_shexp }, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, n_ff_shexp }, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { n_ff_shexp, n_embd }, 0);
layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, mtp_flags);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, n_ff_shexp }, mtp_flags);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, n_ff_shexp }, mtp_flags);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { n_ff_shexp, n_embd }, mtp_flags);
// NextN-specific tensors that define the MTP block.
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, 0);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, 0);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, 0);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, TENSOR_NOT_REQUIRED);
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, mtp_flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, mtp_flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, mtp_flags);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, mtp_flags|TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab }, mtp_flags|TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, mtp_flags|TENSOR_NOT_REQUIRED);
};
for (int i = 0; i < n_layer; ++i) {

6
src/models/step35.cpp
View File

@@ -48,7 +48,11 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
if (!ml.load_mtp) {
mtp_flags |= TENSOR_SKIP;
}
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);