Skip to content

hparams : add n_embd_inp() to support extended embed #16928

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Nov 7, 2025
Merged

hparams : add n_embd_inp() to support extended embed #16928

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0c2aad7
add n_embd_full to support extended embed
CISC Nov 1, 2025
8e28665
don't change output [no ci]
CISC Nov 2, 2025
03341fb
rename to n_embd_inp
CISC Nov 3, 2025
6faa87f
restore n_embd where applicable
CISC Nov 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions include/llama.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -482,6 +482,7 @@ extern "C" {

LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd_full(const struct llama_model * model);
LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model);
Expand Down
14 changes: 7 additions & 7 deletions src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -620,7 +620,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
}

return embd + j*model.hparams.n_embd;
return embd + j*model.hparams.n_embd_full;
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
Expand Down Expand Up @@ -808,7 +808,7 @@ int llama_context::encode(const llama_batch & batch_inp) {

const auto & hparams = model.hparams;

const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_full;
const int64_t n_vocab = model.vocab.n_tokens();

// note: during encode, we always pass the full sequence starting from pos = 0
Expand Down Expand Up @@ -977,7 +977,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;

const int64_t n_vocab = vocab.n_tokens();
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_full;

// when computing embeddings, all tokens are output
const bool output_all = cparams.embeddings;
Expand Down Expand Up @@ -1276,7 +1276,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {

const auto n_batch = cparams.n_batch;
const auto n_vocab = vocab.n_tokens();
const auto n_embd = hparams.n_embd;
const auto n_embd = hparams.n_embd_full;

bool has_logits = true;
bool has_embd = cparams.embeddings;
Expand Down Expand Up @@ -1340,7 +1340,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {

void llama_context::output_reorder() {
const uint64_t n_vocab = model.vocab.n_tokens();
const uint64_t n_embd = model.hparams.n_embd;
const uint64_t n_embd = model.hparams.n_embd_full;

for (size_t s = 0; s < output_swaps.size(); ++s) {
const uint64_t i0 = output_swaps[s].i0;
Expand Down Expand Up @@ -1883,7 +1883,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
{
LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__);

const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd);
const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd_full);

io.write(&embd_size, sizeof(embd_size));

Expand Down Expand Up @@ -2135,7 +2135,7 @@ void llama_context::opt_epoch_iter(
batch.logits [pos_batch] = true;
}

if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_full, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
return;
}
Expand Down
4 changes: 2 additions & 2 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1142,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(

// input embeddings with optional lora
ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_full;

auto inp = std::make_unique<llm_graph_input_embd>();

Expand Down Expand Up @@ -1279,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
// return cur;
//}

const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_full;
const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;

cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
Expand Down
1 change: 1 addition & 0 deletions src/llama-hparams.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@ struct llama_hparams {

uint32_t n_ctx_train; // context size the model was trained on
uint32_t n_embd;
uint32_t n_embd_full; // main + auxiliary embeds
uint32_t n_embd_features = 0;
uint32_t n_layer;
int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
Expand Down
23 changes: 10 additions & 13 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -276,7 +276,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
} break;
case GGML_OP_IM2COL:
{
const int n_embd = hparams.n_embd;
const int n_embd = hparams.n_embd_full;
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
} break;
Expand Down Expand Up @@ -505,6 +505,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
ml.get_key(LLM_KV_EXPERT_GROUP_COUNT, hparams.n_expert_groups, false);
ml.get_key(LLM_KV_EXPERT_GROUP_USED_COUNT, hparams.n_group_used, false);
hparams.n_embd_full = hparams.n_embd;

if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
Expand Down Expand Up @@ -1041,7 +1042,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
hparams.n_embd_full *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_QWEN3MOE:
{
Expand All @@ -1067,7 +1068,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
hparams.n_embd_full *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_PHI2:
{
Expand Down Expand Up @@ -3332,10 +3333,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_QWEN3:
case LLM_ARCH_QWEN3VL:
{
// for model loading, the weights only have the main embd
// so we need to divide by the number of deepstack layers + 1
// n_embd is const int so we declare a new variable
int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

// output
Expand Down Expand Up @@ -3371,10 +3368,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_QWEN3MOE:
case LLM_ARCH_QWEN3VLMOE:
{
// for model loading, the weights only have the main embd
// so we need to divide by the number of deepstack layers + 1
// n_embd is const int so we declare a new variable
int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

// output
Expand Down Expand Up @@ -6681,8 +6674,8 @@ ggml_backend_buffer_type_t llama_model::select_buft(int il) const {
return ::select_buft(
*pimpl->dev_layer.at(il).buft_list,
[&](ggml_context * ctx) {
ggml_tensor * cur = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
ggml_tensor * layer_dir = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
ggml_tensor * cur = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd_full);
ggml_tensor * layer_dir = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd_full);
return ggml_add(ctx, cur, layer_dir);
});
}
Expand Down Expand Up @@ -7329,6 +7322,10 @@ int32_t llama_model_n_embd(const llama_model * model) {
return model->hparams.n_embd;
}

int32_t llama_model_n_embd_full(const llama_model * model) {
return model->hparams.n_embd_full;
}

int32_t llama_model_n_layer(const llama_model * model) {
return model->hparams.n_layer;
}
Expand Down
3 changes: 1 addition & 2 deletions src/models/qwen3vl-moe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,9 +1,8 @@
#include "models.h"

llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;

GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
Expand Down
5 changes: 1 addition & 4 deletions src/models/qwen3vl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,10 @@
#include "models.h"

llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {

const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
const int64_t n_embd = hparams.n_embd;
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);

Expand Down
2 changes: 1 addition & 1 deletion tools/mtmd/mtmd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,7 +151,7 @@ struct mtmd_context {
print_timings(ctx_params.print_timings),
n_threads (ctx_params.n_threads),
media_marker (ctx_params.media_marker),
n_embd_text (llama_model_n_embd(text_model))
n_embd_text (llama_model_n_embd_full(text_model))
{
if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
Expand Down
Loading