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

Open
wants to merge 3 commits into
base: master
Choose a base branch
from
Open

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

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 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
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_inp (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
8 changes: 4 additions & 4 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_inp();
Copy link
Collaborator

@ngxson ngxson Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Technically this should be n_embd_out (aka currently n_embd), right?

Copy link
Collaborator Author

@CISC CISC Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could be, I wasn't entirely sure.

} 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_inp();
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_inp();

// when computing embeddings, all tokens are output
const bool output_all = cparams.embeddings;
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_inp(), 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_inp();

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_inp();
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
10 changes: 10 additions & 0 deletions src/llama-hparams.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
return n_head/n_head_kv;
}

uint32_t llama_hparams::n_embd_inp() const {
uint32_t n_embd_inp = n_embd;

if (n_deepstack_layers > 0) {
n_embd_inp += n_embd * n_deepstack_layers;
}

return n_embd_inp;
}

uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
const uint32_t n_head_kv = this->n_head_kv(il);

Expand Down
3 changes: 3 additions & 0 deletions src/llama-hparams.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,6 +227,9 @@ struct llama_hparams {

uint32_t n_gqa(uint32_t il = 0) const;

// dimension of main + auxiliary input embeddings
uint32_t n_embd_inp() const;

// dimension of key embeddings across all k-v heads
uint32_t n_embd_k_gqa(uint32_t il = 0) const;

Expand Down
28 changes: 10 additions & 18 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -276,8 +276,8 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
} break;
case GGML_OP_IM2COL:
{
const int n_embd = hparams.n_embd;
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
const int n_embd_inp = hparams.n_embd_inp();
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1);
Copy link
Collaborator Author

@CISC CISC Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm unsure if this is correct as well...

Copy link
Collaborator

@ngxson ngxson Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think either n_embd_inp or n_embd should not be very important. at least for now, models that use this ops will have n_embd_inp == n_embd

op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
} break;
case GGML_OP_SCALE:
Expand Down Expand Up @@ -1039,9 +1039,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case 64: type = LLM_TYPE_32B; break;
default: type = LLM_TYPE_UNKNOWN;
}
// 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;
} break;
case LLM_ARCH_QWEN3MOE:
{
Expand All @@ -1065,9 +1062,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case 94: type = LLM_TYPE_235B_A22B; break;
default: type = LLM_TYPE_UNKNOWN;
}
// 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;
} break;
case LLM_ARCH_PHI2:
{
Expand Down Expand Up @@ -3332,10 +3326,6 @@ 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 +3361,6 @@ 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 @@ -6482,6 +6468,7 @@ void llama_model::print_info() const {
if (!hparams.vocab_only) {
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
LLAMA_LOG_INFO("%s: n_embd_inp = %u\n", __func__, hparams.n_embd_inp());
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str());
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
Expand Down Expand Up @@ -6681,8 +6668,9 @@ 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);
const int n_embd_inp = hparams.n_embd_inp();
ggml_tensor * cur = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd_inp);
ggml_tensor * layer_dir = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd_inp);
Comment on lines +6671 to +6673
Copy link
Collaborator

@ngxson ngxson Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not 100% sure if n_embd_inp is correct here.

As far as I can understand, the cur is the output of the current layer. That means it should use the internal embedding length n_embd. CC @65a in case you know this part.

Copy link
Contributor

@65a 65a Nov 4, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So I just know what I learned from trying to make cvectors for Qwen3VL, so with that in mind:

  1. The Transformers hiddenstate (and output embedding, being last_hidden_state) size of that model is 5120 (n_embd) and not 20480 (n_embd_inp). But I didn't use multimodal, and not sure if it changes.
  2. I think the deepstack layers are only used for multimodal

This at least is called from llama-adapter.cpp for cvectors, but the check in cvector is hardcoded against n_embd, so it will still pass regardless (was wondering why my cvector worked with this code change). I assume since the cvector is for the language model, it doesn't matter if the other ones are empty so much so long as they are zeroed.

It seems like the answer might be use n_embd here unless there is somewhere multimodal where this is used, or multimodal cvectors are a thing, or this is called from other contexts that need the "rest" of it?

return ggml_add(ctx, cur, layer_dir);
});
}
Expand Down Expand Up @@ -7329,6 +7317,10 @@ int32_t llama_model_n_embd(const llama_model * model) {
return model->hparams.n_embd;
}

int32_t llama_model_n_embd_inp(const llama_model * model) {
return model->hparams.n_embd_inp();
}

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_inp(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