Skip to content

model: (qwen3next) correct vectorized key_gdiff calculation #19324

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
ngxson merged 2 commits into from
Feb 4, 2026
Merged

model: (qwen3next) correct vectorized key_gdiff calculation #19324

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4bfbf0b
model: (qwen3next) correct vectorized key_gdiff calculation
ngxson Feb 4, 2026
d871ac8
move transpose to outside of loop
ngxson Feb 4, 2026
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
12 changes: 9 additions & 3 deletions src/models/qwen3next.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -265,9 +265,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)

ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
1, chunk_size, n_chunks, g_diff_exp->ne[3]);

ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)

ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)


// state to be updated per chunk
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
Expand Down Expand Up @@ -322,9 +328,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);

// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk));
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff)));
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);

// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
Expand Down
Loading