Skip to content

Attempt to remove embd mangling for Qwen3VL and Qwen3VL-MOE #16925

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

Closed
wants to merge 3 commits into from
Closed

Attempt to remove embd mangling for Qwen3VL and Qwen3VL-MOE #16925

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
b85cddb
Remove n_embd hack from llama-models.cpp
65a Nov 1, 2025
981d578
Remove embd hack from qwen3vl.cpp
65a Nov 1, 2025
af5f50f
Remove embd hack from qwen3vl-moe.cpp
65a Nov 1, 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
16 changes: 2 additions & 14 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
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 +3326,7 @@ 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);
int64_t n_embd = hparams.n_embd;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

// output
Expand Down Expand Up @@ -3371,10 +3362,7 @@ 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);
int64_t n_embd = hparams.n_embd;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

// output
Expand Down
8 changes: 4 additions & 4 deletions src/models/qwen3vl-moe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
#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_full = hparams.n_embd * (n_deepstack_layers + 1); // main embd + deepstack embds
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 All @@ -21,9 +21,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_

if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd_full, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd_full, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
Expand Down
8 changes: 4 additions & 4 deletions src/models/qwen3vl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,9 @@

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_full = hparams.n_embd * (n_deepstack_layers + 1); // main embd + deepstack embds
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;


Expand All @@ -23,9 +23,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_

if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd_full, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd_full, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
Expand Down
Loading